`

IBM Spectrum Protect

In-the-Cloud Deployment
Guidelines with Amazon Web
Services

Document version 1.2

James Damgar
Daniel Benton
Jason Basler
IBM Spectrum Protect Performance Evaluation

© Copyright International Business Machines Corporation 2017.

US Government Users Restricted Rights – Use, duplication or disclosure restricted by GSA ADP Schedule
Contract with IBM Corp.
© Copyright International Business Machines Corporation 2018, 2019

US Government Users Restricted Rights – Use, duplication or disclosure restricted by GSA ADP Schedule
Contract with IBM Corp.
 CONTENTS

Contents.............................................................................................................................................. 3
List of Figures ...................................................................................................................................... 5
List of Tables ....................................................................................................................................... 6
Introduction ........................................................................................................................................ 6
1.1 Purpose of this Paper.......................................................................................................... 6
1.2 Considerations for Disk-to-Cloud Tiering Versus Direct-to-Cloud Data Movement ........... 7
1.2.1 Cloud Accelerator Cache Considerations.................................................................... 8
1.2.2 Workload Limitations and Considerations with Tiering ............................................. 9
1.2.3 Amazon S3 Intelligent-Tiering ................................................................................... 12
1.3 Cloud Deployment Patterns.............................................................................................. 12
1.4 Cloud Environment Considerations .................................................................................. 14
1.4.1 Importance of Adequate Sizing ................................................................................ 14
1.4.2 Linux Logical Volume Manager (LVM) ...................................................................... 15
1.5 References to Physical IBM Spectrum Protect Blueprints ................................................ 15
1.6 Database Backup Capacity ................................................................................................ 16
1.7 Server Maintenance Scheduling Considerations .............................................................. 17
1.8 Session Scalability by Blueprint Size ................................................................................. 17
Amazon EC2 Compute Configurations.............................................................................................. 19
2.1 Design Considerations for Amazon EC2 Instances ........................................................... 26
2.1.1 Considerations for Direct-to-Cloud Architectures .................................................... 28
2.1.2 Sizing the Cloud Accelerator Cache .......................................................................... 29
2.1.3 AWS: Large Instance Considerations ........................................................................ 30
2.1.4 AWS: Medium Instance Considerations ................................................................... 31
2.1.5 AWS: Small Instance Considerations ........................................................................ 31
2.1.6 AWS: Extra-Small Instance Considerations ............................................................... 31
Throughput Measurements and Results .......................................................................................... 32
3.1 Dataset Descriptions ......................................................................................................... 32
3.2 Backup and Restore Measurements ................................................................................. 34
3.2.1 AWS: Large Instance Measurements ........................................................................ 34
3.2.2 AWS: Medium Instance Measurements ................................................................... 37
3.2.3 AWS: Small Instance Measurements ........................................................................ 39
3.2.4 AWS: Extra-Small Instance Measurements .............................................................. 40
Appendix ........................................................................................................................................... 42
Disk Setup Commands for Linux Deployments ................................................................................. 42
Disk Benchmarking ........................................................................................................................... 57
Object Storage Benchmarking .......................................................................................................... 64
Instance and Object Storage: Navigating the AWS Portal ................................................................ 68
References ........................................................................................................................................ 74
Notices .............................................................................................................................................. 75
Trademarks ........................................................................................................................... 76
 LIST OF FIGURES

Figure 1: Disk-to-cloud tiering, before and after .............................................................................. 10

Figure 2: Disk-to-cloud tiering, after tiering ...................................................................................... 11

Figure 3: Deployment patterns ........................................................................................................ 13

Figure 4: Sizing the cloud accelerator cache for AWS .................................................................... 30

Figure 5: AWS large configuration; database volume average throughput; 8 KiByte random
writes/reads .................................................................................................................. 60

Figure 6: AWS large configuration; database volume average IOPS; 8 KiByte random writes/reads
...................................................................................................................................... 61

Figure 7: AWS large configuration; cloud cache volume average throughput; mixed 256 KiByte
writes and reads ........................................................................................................... 61

Figure 8: AWS medium configuration; database volume average throughput; 8 KiByte random
writes/reads .................................................................................................................. 62

Figure 9: AWS medium configuration; database volume average IOPS; 8 KiByte random
writes/reads .................................................................................................................. 62

Figure 10: AWS medium configuration; cloud cache volume average throughput; mixed 256 KiByte
writes and reads ........................................................................................................... 63

Figure 11: AWS small configuration; database volume average throughput; 8 KiByte random
writes/reads .................................................................................................................. 63

Figure 12: AWS small configuration; database volume average IOPS; 8 KiByte random
writes/reads .................................................................................................................. 64

Figure 13: AWS small configuration; cloud cache volume average throughput; mixed 256 KiByte
writes/reads .................................................................................................................. 64
 LIST OF TABLES

Table 1: IBM Spectrum Protect physical Blueprint targets (V4.1, Linux x86) .................................. 16

Table 2: Preferred ranges of maximum values for client session counts ........................................ 18

Table 3: AWS, large configuration ................................................................................................... 20

Table 4: AWS, medium configuration .............................................................................................. 21

Table 5: AWS, small configuration ................................................................................................... 23

Table 6: AWS, extra-small configuration ......................................................................................... 24

Table 7: AWS instance options ........................................................................................................ 27

Table 8: Throughput measurement datasets ................................................................................... 32

Table 9: AWS, large configuration, 128 MiByte VE-like dataset backup results ............................. 34

Table 10: AWS, large configuration, 128 MiByte VE-like dataset restore ....................................... 35

Table 11: AWS, large configuration, 128 MiByte random dataset backup results .......................... 35

Table 12: AWS, large configuration, 128 MiByte random dataset restore ...................................... 35

Table 13: AWS, large configuration, 1 GiByte dataset backup results ............................................ 36

Table 14: AWS, large configuration, 1 GiByte dataset restore results ............................................ 36

Table 15: AWS, medium configuration, 128 MiByte VE-like dataset backup results ...................... 37

Table 16: AWS, medium configuration, 128 MiByte VE-like dataset restore results ....................... 37

Table 17: AWS, medium configuration, 128 MiByte random dataset backup results...................... 38

Table 18: AWS, medium configuration, 128 MiByte random dataset restore results ...................... 38

Table 19: AWS, medium configuration, 1 GiByte dataset backup results ....................................... 38

Table 20: AWS, medium configuration, 1 GiByte dataset restore results ....................................... 38

Table 21: AWS, small configuration, 128 MiByte VE-like dataset backup results ........................... 39

Table 22: AWS, small configuration, 128 MiByte VE-like dataset restore results ........................... 39

Table 23: AWS, small configuration, 128 MiByte random dataset backup results .......................... 39

Table 24: AWS, small configuration, 128 MiByte random dataset restore results .......................... 40

Table 25: AWS, small configuration, 1 GiByte dataset backup results ........................................... 40

Table 26: AWS, small configuration, 1 GiByte dataset restore results ............................................ 40

Table 27: AWS, extra-small configuration, 128 MiByte VE-like dataset backup results ................. 40

Table 28: AWS, extra-small configuration, 128 MiByte VE-like dataset restore results .................. 41

Table 29: AWS, extra-small configuration, 128 MiByte random dataset backup results ................. 41
Table 30: AWS, extra-small configuration, 128 MiByte random dataset restore results ................. 41

Table 31: AWS, extra-small configuration, 1 GiByte dataset backup results .................................. 41

Table 32: AWS, extra-small configuration, 1 GiByte dataset restore results ................................... 41
 Introduction

1.1 Purpose of this Paper

This document introduces possibilities for integrating an IBM Spectrum Protect server with
an Amazon Web Services (AWS) cloud computing system. You can use the
configurations as starting points to deploy a large, medium, or small system (as defined in
the IBM Spectrum Protect Blueprints) or an extra-small system. With the goal of achieving
a target daily ingestion rate (corresponding to a large, medium, small, or extra-small
deployment), configuration possibilities are offered so that you can get a sense of the
relative CPU, memory, disk, and network capabilities that are needed to satisfy
requirements. In addition, a survey of options for fulfilling these needs is provided. Certain
cloud instance and disk types offered by providers might be sufficient in some areas while
lacking in others. You must recognize where system bottlenecks might arise that could limit
IBM Spectrum Protect capability.
Use this paper as a starting point for guidance about where to deploy an instance of the
IBM Spectrum Protect server within an Amazon Elastic Compute Cloud (Amazon EC2)
dedicated or shared compute instance with the goal of eventually storing the bulk of
primary backup and archive data on cost-effective object storage. In the case of Amazon,
this means Amazon S3 (Simple Cloud Storage Service). This goal can be accomplished by
configuring an IBM Spectrum Protect cloud-container storage pool with (block disk-based)
accelerator cache. Two approaches are generally available for storing data in an IBM
Spectrum Protect cloud-container storage pool: a direct-to-cloud approach or a disk-to-
cloud tiering approach.
For direct-to-cloud architectures, backup data is ingested directly into a cloud-container
storage pool with a performant accelerator cache disk location tuned for a system’s
ingestion workload as the initial “landing spot” (for more information, see Sizing the Cloud
Accelerator Cache). Data is then immediately, asynchronously transferred to object
storage while further data is also ingested into the disk staging area (also known as
overlapped I/O). The key consideration here is to determine the performance
characteristics that the disk staging area must provide to allow for this mixed write-and-
read behavior to ensure that ingestion targets are met. A Cloud Cache and Object Storage
Benchmarking guide and “Cloud benchmarking tools” packages are provided along with
this paper to assist in benchmarking both the cloud accelerator cache and object storage
system from a prospective host server.
 In contrast, disk-to-cloud tiering architectures make use of IBM Spectrum Protect storage
rules to demote data from one or more (usually small) directory-container storage pools (a
disk tier) to a cloud-container storage pool. Backup data is initially ingested into a directory-
container storage pool and later a portion of this data is moved asynchronously to a cloud-
container storage pool. This can be done with either age-based tiering (available as of IBM
Spectrum Protect Version 8.1.3) or tiering by backup state (available as of IBM Spectrum
Protect V8.1.6). With IBM Spectrum Protect V8.1.6, a combination of storage rules and
storage subrules can be used to facilitate more granular tiering behavior between the disk
and object storage tiers, allowing for flexibility and filtering by node and node file space, for
example. Some guidance is given in this paper to assist with determining whether tiering is
suitable for your workload demands and characteristics. (For an introduction to the tiering
features that are available in IBM Spectrum Protect, see Tiering data to cloud or tape
storage in the online product documentation.)
With AWS, solutions involving disk-to-cloud tiering can be cost prohibitive due to the higher
cost of Amazon Elastic Block Store (EBS) disk, which might have to be provisioned to
support the disk tier, in contrast to the relatively lower cost of Amazon Simple Cloud
Storage Service (S3) object storage. However, restore requirements of a solution may be
such that having operational recovery data on a fast-performing disk tier is worth this
additional cost. Further guidance is provided in the following section regarding
considerations for using a direct-to-cloud or tiering approach. The architectures referenced
within this paper use a direct-to-cloud approach but can be adjusted to conform to the
requirements of a tiering architecture.

1.2 Considerations for Disk-to-Cloud Tiering Versus Direct-to-

Cloud Data Movement
The primary advantage of the tiering model is that operational recovery data can be
preserved on a localized, fast disk tier for rapid recovery while older copies of data or data
intended for long-term retention can be demoted to object storage, which is typically more
affordable. The tiering model can also be used as an alternative to the direct-to-cloud
model with a relatively small disk tier footprint (not strictly for operational recovery
purposes). When the TIERDELAY parameter is set to 0, age-based tiering can be used to
tier each day’s worth of ingested client data after it is ingested (after the backup operation
is completed). In this case, potentially less expensive disk can be provisioned for use by
the small disk container pool tier because no ingest and cloud transfer input/output (I/O)
operations occur in parallel. Tiering can be run serially after the completion of data
ingestion during scheduled windows with less or no contention for this disk; the disk area
can be cleared in preparation for the next day’s ingestion.
The same ingestion targets can be satisfied with the disk-to-cloud tiering model as with the
direct-to-cloud model, assuming that the direct-to-cloud approach makes use of an
accelerator cache and overlapped data ingestion.
Restriction: To implement cloud tiering, you must provision enough disk space to hold a
full day’s worth of ingested data (plus some buffer) to avoid failed backup operations. The
same underlying disk technology can be used in both cases. If, however, you plan to use
disk-to-cloud tiering to hold one or more days’ worth of operational recovery data within a

7
container pool disk tier, the instance disk capacity might have to be much greater, with the
caveat that a slower-performing disk might be sufficient for this case. In all cases, you must
understand the ingestion targets (after data deduplication and compression) to determine a
daily disk capacity for a transient disk case. Meanwhile, operational recovery requirements
in terms of the number of days’ worth of recovery data (after deduplication and
compression) should be determined to further size a disk container pool with tiering to
cloud if necessary.
With the direct-to-cloud model, you can minimize local block storage capacity. This is an
advantage because local block storage can be cost prohibitive in cloud-hosted
environments.

1.2.1 Cloud Accelerator Cache Considerations

Beginning with IBM Spectrum Protect V8.1.2, data ingestion from clients is throttled if the
accelerator cache area is near capacity. This feature makes it possible for this disk cache
location to be underprovisioned from a capacity standpoint in that the disk cache location
does not have to be sized large enough to hold a full day’s worth of deduplicated and
compressed data. However, the accelerator disk still must be performant enough in terms
of input/output operations per second (IOPS) so that client data ingestion and replication
target activity can be completed in a timely manner. In the end, you have to compare costs
to determine whether larger capacity, less-expensive disk with tiering has an advantage
over a direct-to-cloud cache model for a given environment, ingestion rate, and recovery
objective.
Restriction: If you plan to use the direct-to-cloud ingestion model, the cloud accelerator
cache should be sized large enough to hold at least two times the largest front-end object
being ingested. For example, if a 512 GB object is to be ingested directly into a cloud-
container storage pool, the cloud accelerator cache should be at least 1 TB in size.
Similarly, if 5 client sessions will be backing up 100 GB files each at the same time, the
cloud accelerator cache should be sized to at least 1000 GB (5 clients x 100 GB files x 2).
This is because the IBM Spectrum Protect server will attempt to “reserve” space in the
cloud accelerator cache for in-flight ingestion until ingestion is completed for those objects
and their database transactions are committed to the server. By default, this processing
assumes no deduplication or compression savings and attempts to reserve the total front-
end amount of data to ensure sufficient storage capacity.
Beginning with IBM Spectrum Protect V8.1.6, a server option can be used to influence this
behavior. The PreallocReductionRate server option can be used to give the server a
“hint” about the expected reduction ratio for ingested data and cause the server to reserve
less physical space in the container storage pool. For example, setting this option to 5 will
cause the server to assume a 5:1 data reduction rate for front-end to back-end data so that
only 1 unit of back-end space will be reserved for 5 units of front-end protected data. This
option can range from 1 (the default, no reduction) to 8 (an 8:1 assumed reduction). Use
this option only when a smaller cloud accelerator cache is desired and data reduction rates
are certain. If the storage pool has inadequate space, backup failures can occur.
1.2.2 Workload Limitations and Considerations with Tiering
Not all client workloads are suitable for a disk-to-cloud tiering model. Tiering by age (as of
IBM Spectrum Protect V8.1.3) allows for the demotion of backup objects that have reached
the specified age threshold. Inactive backup generations that are older than the specified
age are transitioned to object storage. Tiering by state (as of IBM Spectrum Protect V8.1.6)
allows for the demotion of backup objects that have reached the specified age threshold
and are inactive within the server. Active backup objects are preserved on the disk tier,
while inactive copies of backup objects are transitioned to object storage.
Disk-to-cloud tiering is suitable for client workloads that have low data deduplication rates
(backup generations differ greatly). In this case, data is highly unique between backup
operations. When a backup generation is tiered to object storage, the deduplicated extents
(chunks) that make up that object have their references decremented on the source
directory-container storage pool. In this case, reference counts are likely to be low and
more deduplicated extents are likely to be removed from the disk tier as objects are tiered
and space is released.
Disk-to-cloud tiering might not be suitable for client workloads that have a high data
deduplication rate. In this case, data is not very unique between backup generations and
many shared deduplicated extents are referenced by multiple object generations. Even
though an object can be tiered by a storage tiering rule, because the object shares many
extents with other objects (which might still be active), a large proportion of the object’s
data will not be removed from the disk tier (although it will be copied to the object storage
tier).
The following figures illustrate how data movement with disk-to-cloud tiering can occur.
Figure 1 depicts a scenario in which multiple versions of three backup objects (A, B, and C)
have been ingested and are stored in a directory-container storage pool on disk. Dotted
lines represent references to deduplicated extents (colored, numbered boxes). With the
tier-by-state option, the inactive object copies (shown in the gray rectangle) would be tiered
to a cloud-container storage pool.

9
Figure 1: Disk-to-cloud tiering, before and after

Figure 2 depicts the situation after tiering is completed and the REUSEDELAY parameter
value of the source directory-container storage pool is exceeded (so that deduplicated
extent removal for extents with zero reference count can occur).
Figure 2: Disk-to-cloud tiering, after tiering

Notice that deduplicated extents 1 and 2 remain on disk even after tiering and extent
cleanup have occurred. This is due to the fact that those extents are shared between the
active and inactive backup copies. If many deduplicated extents are shared by objects (a
high duplicate data rate with high data deduplication ratios), it is more likely that data will
remain on disk, even after backup objects have been tiered at an IBM Spectrum Protect
inventory level. Keep this factor in mind when you consider a disk-to-cloud tiering model
and when you size an environment.
For workloads that deduplicate well from day to day, there will be many shared extents
across backup and archive generations and a smaller capacity footprint on tiered object
storage as a result because these backup and archive generations will also share many
extents in the cloud-container storage pool. For workloads that deduplicate poorly day to
day (highly unique data change each day), there will be few shared extents across backup
and archive generations and potentially a larger capacity footprint on tiered object storage
because these backup and archive generations will each point to (more) unique data in the
cloud-container storage pool.
If the primary motivation for using disk-to-cloud tiering is rapid recovery of operational data,
a tiering model might provide the best approach. You must understand the nature of the
client workload to accurately size the directory-container storage pool on disk.

11
1.2.3 Amazon S3 Intelligent-Tiering
Beginning with IBM Spectrum Protect V8.1.8, you can enable Amazon S3 Intelligent-
Tiering for objects in IBM Spectrum Protect cloud-container storage pools. When you issue
the DEFINE STGPOOL command to define a cloud-container storage pool with Amazon
S3, you can specify a new parameter, CLOUDSTORAGECLASS (and update the parameter
value later by using the UPDATE STGPOOL command). Setting this parameter to a value of
AUTOMATICVENDORTIERING will cause all new cloud-container storage pool objects that
are uploaded to S3 to be sent to the Intelligent-Tiering storage class (as opposed to the
default Standard storage class). Objects with the AUTOMATICVENDORTIERING storage
class type can be transitioned automatically to the Amazon S3 Infrequent Access tier if the
object has not been accessed for 30 consecutive days. This functionality offers storage
capacity savings for Amazon S3 users. IBM Spectrum Protect data that is already
deduplicated and compressed can be further transitioned to a lower-cost storage tier within
S3, provided that restore or other read requests that would involve deduplicated extents
within the transitioned object are infrequent. With Amazon S3 Intelligent-Tiering, objects
that are read are automatically moved back to the frequent access tier. For more
information about Amazon Intelligent-Tiering, see References [10].

1.3 Cloud Deployment Patterns

The described configurations can be used as starting points in situations where the IBM
Spectrum Protect cloud instance will be a primary server and in situations where it is
used as a replication target. In scenarios where the cloud-based instance is a replication
target, adequate “public” network capability might be necessary to satisfy replication
throughput requirements. AWS Direct Connect can be used to establish a dedicated 1
Gbps or 10 Gbps network connection from an on-premises data center to AWS to facilitate
efficient IBM Spectrum Protect replication or backup processing from peer servers or
clients outside the AWS infrastructure.
Generally, IBM Spectrum Protect deployments making use of cloud-based object storage
will align with one of the following three patterns:
Figure 3: Deployment patterns

In the figure, the first deployment pattern could involve an IBM Spectrum Protect server
that is installed on premises or in Amazon on an EC2 instance, with primary backup and
archive data landing in object storage immediately. The positioning of the IBM Spectrum
Protect server in relationship to clients could be one critical decision point when you
consider whether to have a server instance on premises or within AWS. This pattern could
involve use of a direct-to-cloud architecture with accelerator cache or a small disk
container pool with immediate tiering to a second cloud-container storage pool without
accelerator cache.
The second deployment pattern would make use of cloud-based Amazon S3 object
storage at the secondary disaster recovery (DR) site. This DR server could be installed at
an on-premises site or on an Amazon EC2 instance. In the latter case, sufficient wide area
network (WAN) bandwidth between the primary and secondary sites is required for
acceptable performance. Much like the first deployment pattern, here the IBM Spectrum
Protect server at the DR site could make use of a direct-to-cloud topology with a cloud-
container storage pool featuring accelerator cache, or it could use a small disk container
pool landing spot with immediate tiering to a cloud-container storage pool backed by object
storage.
The third deployment pattern features specific use of disk-to-cloud tiering, available with
IBM Spectrum Protect V8.1.3 and later, to allow for operational recovery data to reside on
faster performing disk storage. Data that is older, archived, or both would be tiered to
cloud-based Amazon S3 object storage after a specified number of days. This deployment
could also be implemented at an on-premises site or within an Amazon EC2 instance.
However, the additional cost of having a larger capacity disk container pool should be
factored into cost estimates with an in-the-cloud solution.
A combination of approaches is also possible within the same deployment. For example,
a cloud-container storage pool could be configured with accelerator cache disk and made
to store long-term retention or compliance archives. A directory-container storage pool
could be configured as a disk tier for normal backups, and a tiering relationship could be
set up so that operational recovery data (for example, backups from the previous 7 days) is
kept on this disk tier, while older data is demoted to the same cloud-container storage pool.

13
 The same cloud-container storage pool can be a direct backup target and a tiering target.
However, if the pool is a direct target of a backup-archive client, the pool must be
configured with accelerator cache disk.

1.4 Cloud Environment Considerations

For Amazon EC2, you can choose one of multiple potentially suitable instance types to
satisfy an IBM Spectrum Protect server requirement. You do not have complete flexibility in
terms of the type and proportion of CPU and memory resources available for an instance.
However, standard instance types are available, which can roughly match recommended
server requirements. Take care to select instances that have the required CPU and
memory resources to support inline server data deduplication, compression, encryption,
and cloud application programming interface (API) operations. The instances also must
have the associated storage area network (SAN) disk (Amazon EBS) and network
(Ethernet) capability to support the desired workload.
With Amazon Elastic Cloud Compute (EC2) instances, you can alter the instance type
later, when resource requirements might be clearer. The process is fairly straightforward
for Elastic Block Store (EBS) backed instances using Amazon’s web portal. However, there
are some caveats to keep in mind. For details, see the Amazon documentation under
References [3]. In particular, experience shows that driver issues can occur when
attempting to resize Windows-based EC2 instances after initial provisioning. To help avoid
complications, use Red Hat Enterprise Linux or Ubuntu Linux instances in Amazon EC2
with IBM Spectrum Protect.

1.4.1 Importance of Adequate Sizing

Ingested backup data reaching cloud accelerator cache or the initial disk container pool tier
requires the use of block storage allocated to the cloud server. IBM Spectrum Protect
database activity also uses some level of throughput and elevated I/O operations during
workload processing. Therefore, disk I/O capability and instance-to-disk throughput
considerations must be evaluated when choosing and configuring an in-the-cloud IBM
Spectrum Protect instance. With a final ingestion point on object storage via a cloud-
container storage pool, the Ethernet capability of the instance and the nature of the
network between the instance and the Amazon S3 object storage endpoint must be kept in
mind. As part of this, consider how the front-end client data workload deduplicates and
compresses data into the back-end stored quantity.
Certain Amazon instances might or might not have access to dedicated bandwidth to
attached EBS disks. A lack of access can create a bottleneck in the IBM Spectrum Protect
server’s database operations. Certain instances might have limited throughput over
Ethernet, and this limitation could hamper ingestion and restore throughput with object
storage. During the planning phase, consider how ingested data will be reduced by data
deduplication and compression when the data moves to the back-end storage location.
These factors will help you estimate how much back-end data must be moved within a
certain time window (measured in hours) and can help predict the throughput (megabytes
per second or terabytes per hour) that the Ethernet network and object storage endpoint
require to satisfy ingestion requirements. Generally, 10 Gbps Ethernet capability to private
 Amazon S3 storage endpoints is required for large, medium, or small Blueprint ingestion
targets, while 1 Gbps is sufficient for extra-small targets.
Beginning with IBM Spectrum Protect V8.1.3, the server automatically throttles client
backup operations if the cloud accelerator cache portion of a cloud-container storage pool
is nearing full capacity. As a result, it is not mandatory to configure cloud accelerator disk
cache space that would be large enough to hold a full day’s worth of backups (after data
deduplication and compression). However, disk benchmarks should be run to ensure that
the anticipated back-end workload that an IBM Spectrum Protect server is expected to
support will not result in this disk location being the primary bottleneck of the system (see
Disk Benchmarking). In practice, any planned deployment should be validated to ensure
that it will meet performance requirements.

1.4.2 Linux Logical Volume Manager (LVM)

The described reference architectures use either the Red Hat Enterprise Linux or Ubuntu
Linux operating system. If you want to deploy a solution on Linux based operating systems,
the preferred method is to use the Linux Logical Volume Manager (LVM) for the cloud
accelerator cache disk and, optionally, the IBM Spectrum Protect database archive log disk
(when more than one physical disk is utilized). The remaining IBM Spectrum Protect disk
components can be satisfied with file systems formatted on directly mounted Amazon EBS
block disks. The overlapped I/O pattern experienced with concurrent backup ingestion
activity and transfer of data to object storage can lead to “hot spots” on disk when more
than one storage pool directory is defined for a cloud-container storage pool as accelerator
cache. To help avoid a throughput bottleneck, you can configure a single logical volume to
span all physical volumes assigned for use as cloud accelerator cache. Furthermore, the
preferred method is to use a stripe size of 16 KiBytes for the single logical volume and
ensure that the number of stripes matches the number of physical disks. For more
information, see Disk Setup Commands for Linux Deployments.
The use of Linux LVM to logically stripe across several Amazon EBS disks should not be
depended upon to increase the durability of the underlying storage, as Amazon EBS disks
are already redundant within an Amazon Availability Zone and so do not benefit from the
recovery characteristics of LVM striping or RAID.

1.5 References to Physical IBM Spectrum Protect Blueprints

Throughout this paper, the physical server specifications contained in the latest (V4.1) IBM
Spectrum Protect Blueprint and Server Automated Configuration for Linux x86
document (also known as the IBM Spectrum Protect Blueprints) are referenced as targets
for CPU and memory configurations matching small, medium, and large server builds. For
more information about the Blueprints, see References [1]. The intention with the server
builds outlined here is to provide systems capable enough from a CPU, memory, disk, and
Ethernet point of view to approach Blueprint-level ingestion capability. Although different
instance types can be used to satisfy the same requirements, and deviations from the
Blueprint specifications mentioned in this document are possible, the disk specifications
should be followed to ensure appropriate ingest and database scalability and performance.
As a reference, the following table indicates the throughput, capacity, CPU, and memory
targets for each of the referenced Blueprints. The values for total managed data and daily

15
 ingest data are for the block storage Blueprints. These ingestion targets assume an 8-hour
backup window.
Table 1: IBM Spectrum Protect physical Blueprint targets (V4.1, Linux x86)

Sizing CPU RAM Total managed Daily ingest data

category memory data (front end) (front end)

Small 16 cores 64 GB 60 TB – 240 TB Up to 10 TB per day

Medium 20 cores 128 GB 360 TB – 1440 TB 10 – 30 TB per day

Large 44 cores 384 GB 1000 TB – 4000 TB 20 – 100 TB per day

Although not defined explicitly in the physical Blueprints, the extra-small cloud Blueprint
systems target up to 10 TB or more of total managed (front-end) data with a daily ingestion
rate of up to 1 TB, or more, per day.

1.6 Database Backup Capacity

The configurations in this document are provisioned with enough IBM Spectrum Protect
database backup disk space to hold two days’ worth of full database backups in the worst
case (an IBM Db2 database consuming close to its resident capacity). This approach was
taken to avoid the excessive attached disk cost associated with provisioning excess cloud
disk space for storing many days’ worth of backups. For example, with Amazon EBS, GP2
type volumes cost $0.10 per GiByte-month in the US West (Oregon) Region based on
provisioned space. Compare this with the $0.023 per GiByte-month in the same region for
S3 object storage where pricing is based on actual storage used. Not only is unused
provisioned disk space a deterrent to cost savings, the actual rate charged for this space is
much more than object storage considering that the data involved (database backups) is
archive-like in nature.
The IBM Spectrum Protect server is presently incapable of storing database backups
directly to object storage. For those desiring recovery protection for a longer period (for
example, 7 – 10 days), the database backup disk space outlined here might have to be
increased. Alternatively, a strategy can be employed outside of IBM Spectrum Protect
whereby the disks that underlie the database backup disk can be snapshotted at the cloud
service provider’s level and stored to object storage after a database backup is completed.
This may be driven manually or, more likely, in a scripted fashion by using cloud provider
APIs from within the instance itself. This has been done with success in IBM GTS with
Amazon based IBM Spectrum Protect deployments.
If you use this method, the disk space within the database backup disk location could be
cleared in preparation for the next day’s database backup, allowing for only two days’
worth of footprint on provisioned cloud block disk. The reason that two days’ worth of
capacity is required and not one is that it is impossible to delete the “only” database
backup copy until after a new one is created. These object storage-based disk snapshots
would need to be handled with great care and security. If you want to recover a database
 to a point N days in the past, the disk snapshots for the database disk volumes would have
to be restored from object storage before normal IBM Spectrum Protect database backup
recovery could commence.

1.7 Server Maintenance Scheduling Considerations

The IBM Spectrum Protect Blueprint and Server Automated Configuration for Linux x86
V4.1 document provides a detailed breakdown of the procedure for setting up IBM
Spectrum Protect server maintenance schedules (see References [1], Chapter 5). Use this
information as a reference for establishing a maintenance schedule on cloud-hosted
servers.
For an IBM Spectrum Protect server in AWS that is serving as a replication target, a
replication window and schedule might have to be established. For servers using the
direct-to-cloud model, where primary backup data is ingested directly into a cloud-container
storage pool, a replication window might not be required if this server is not a replication
target server because a cloud-container storage pool cannot be used as a replication
source. In this case, redundancy requirements for the ingested client data can be met by
the inherit redundancy of Amazon S3 object storage.
For an IBM Spectrum Protect server in AWS that is using the disk-to-cloud tiering model, a
replication source strategy might be required. Replication can help to protect client data
objects in the disk directory-container storage pool that have not yet been tiered (demoted)
to object storage because only one copy of that data is present. To prevent excess data
from being stored (pinned) to the disk tier, verify the following items:

• The source replication server (used for disk-to-cloud tiering) should be

configured with a longer retention policy than the target replication server.
In other words, data should be retained for a longer period on the source
replication server.

• The retention policy that affects client node data on the target replication
server should match the value of the TIERDELAY parameter of the storage
rule responsible for tiering the same client node data on the source server.
In general, the server that is used for disk-to-cloud tiering (whether it be the source
replication server or the target replication server) should be the server with the longer
retention policy for the client nodes that are affected by the tiering storage rule.

1.8 Session Scalability by Blueprint Size

The IBM Spectrum Protect Blueprint and Server Automated Configuration for Linux
x86 document describes how to set the IBM Spectrum Protect server option
MAXSESSIONS, based on Blueprint system size:

- Small system: 250 maximum simultaneous client sessions

- Medium system: 500 maximum simultaneous client sessions
- Large system: 1000 maximum simultaneous client sessions

17
(For more information about the Blueprint configurations, see References [1].)
The actual throughput scalability of a cloud-based solution depends on many factors,
including the configured disk capability and capacity of the system, the amount of CPU and
memory resources available on the system, and the relative rate of data deduplication and
compression for the dataset that is ingested into the server. Larger objects, which feature a
larger deduplicated extent size (for example, 250 - 350 KiBytes, or more) and which do not
deduplicate or compress well (for example, less than 10%), will result in less database and
computation (CPU) overhead, but will utilize more disk and network bandwidth. The logical
reduction of front-end client data to the physical back-end data (which is actually written
out and stored to disk and object storage) means that the disk, network, and object storage
components will be stressed to a higher degree as client/server session counts increase.
Memory usage by the IBM Spectrum Protect server might also be greater. As session
counts increase, these components are likely to become a system bottleneck, limiting front-
end throughput.
Objects that feature smaller, deduplicated extent sizes (for example, 60 - 100 KiBytes or
similar) and that deduplicate and compress well (for example, 50% data deduplication with
50% compressibility) will result in less network, disk, and object storage bandwidth used,
but will lead to more database and computation overhead to facilitate these data reduction
operations. As session counts increase, CPU and database-related memory are likely to
first become limiting factors for these data types. In general, the more successfully data
can be deduplicated and compressed (and therefore the greater the data reduction from
front-end to back-end data), the greater the number of feasible client sessions. The
following table indicates a reasonable range of client session counts based on system size
and data type, as well as the likely limiting factor for the system as the high end of the
range is approached. For more information about these data types, see Throughput
Measurements and Results.
Table 2: Preferred ranges of maximum values for client session counts

Cloud Large object, Large object, Large object, Small object,

system poor data good data small extent size, poor data
size deduplication deduplication good data deduplication
and and compression2 deduplication and
compression1 and compression4
compression3

Extra 10 – 50 25 – 50 10 – 50 10 – 50
small

Small 50 - 100 100 - 200 50 - 100 50 - 100

Medium 100 - 200 200 - 400 100 - 150 100 - 150

Large 300 - 400 400 - 500 150 - 200 150 - 200

 Limiting Network, disk, CPU, memory or CPU, memory CPU, memory
factor at object storage network, disk, object
scale bandwidth, storage bandwidth
memory

1
This model uses 128 MiByte objects, 250 - 350 KiByte extents, and <10% data deduplication and
compressibility. Full backup operations are used with pseudo random data or data that cannot be easily
deduplicated or compressed. For example, this model can be applied to encrypted data.
2
This model uses 128 MiByte objects, 150 - 200 KiByte extents, and 50% data deduplication and compressibility.
For example, this model can be applied to virtual machine backups.
3
This model uses 1 GiByte objects, 60 - 100 KiByte extents, and 50% data deduplication and compressibility. For
example, this model can be applied to database image backups.
4
This model uses 128 KiByte objects and <10% data deduplication and compressibility. For example, this model
can be applied to file server data and other small files or objects.

Often, a diminishing rate of return in regard to throughput is experienced when 50 - 100 total client
sessions are exceeded, regardless of data type. More sessions are possible and might be
warranted, given the client schedule or other requirements. However, aggregate gains in total
throughput of a single IBM Spectrum Protect instance might not be substantial past this point.

Amazon EC2 Compute Configurations

For an overview of this Amazon web service, see Amazon EC2.
The following configurations represent preferred IBM Spectrum Protect deployments within
the Amazon Web Services infrastructure. These deployments make use of Amazon
Virtual Private Clouds (VPCs) within a desired Amazon region with a VPC endpoint
configured for the Amazon S3 service. The Amazon S3 bucket utilized by each instance for
IBM Spectrum Protect usage is created within the same Amazon region and is accessed
by the IBM Spectrum Protect server via HTTPS. Amazon guidelines call for creating a VPC
endpoint for the S3 service so that instances accessing the public endpoint address (for
example, using an IBM Spectrum Protect cloud-container storage pool with a cloud URL of
https://s3.us-west-2.amazonaws.com for US West Oregon) will have traffic routed through
Amazon’s internal network rather than being exposed to a broader WAN. This approach
benefits both the reliability and latency of IBM Spectrum Protect cloud operations.
For circumstances in which the IBM Spectrum Protect server is provisioned within an
Amazon EC2 compute instance and this server is the target server for storage pool
protection and node replication operations from a source server located at an on-premises
data center, additional network bandwidth might be needed to satisfy replication
requirements. Consider using AWS Direct Connect to establish a dedicated, high
bandwidth link from your premises to the AWS-based IBM Spectrum Protect instance
running in EC2. Consider this option if your anticipated back-end replication workload will
cause the needed throughput to exceed the bandwidth capacity between the replication
pair. Amazon Direct Connect can be used to establish a dedicated 1 Gbps or 10 Gbps

19
network connection from an on-premises data center to AWS in order to access Amazon
Virtual Private Cloud (VPC) as well as public AWS services, such as Amazon S3.
For example, if an IBM Spectrum Protect source server were to ingest 40 TB of front-end
data that reduced to 10 TB after (2:1) data deduplication and (2:1) data compression, that
10 TB of back-end data would have to be replicated during the daily storage pool
protection and node replication window. A 1 Gbps dedicated link can support
approximately 3.5 TB per hour of throughput. Amazon Direct Connect offers both 1 Gbps
and 10 Gbps options with the additional benefit of being charged an AWS Direct Connect
data transfer rate that is below the rate charged for incoming internet data transfer. See
References [6].For each of the instances, isolated throughput measurements should be
conducted with an S3 API tool to determine maximum throughput to collocated S3 bucket
storage within that region.
Table 3: AWS, large configuration

Cloud component AWS component Detailed Quantity

description

Server and m5.16xlarge or 64-core Intel 1

network m5a.16xlarge Amazon Platinum 8175 @
EC2 instance 3.10GHz or AMD
(dedicated) EPYC 7000 @
2.5GHz

256 GB RAM

EBS disk optimized

with 10000 Mbps
EBS throughput

20 Gbit Ethernet
connectivity

Virtual Private Cloud Instance placed in 1

(VPC) VPC; endpoint
created for
Amazon S3 service
in the same AWS
region

Operating system Red Hat Enterprise 1

Linux or Ubuntu
Linux server

Block storage 100 GB EBS GP2 Operating system 1

volume disk
 100 GB EBS GP2 IBM Spectrum 1
Protect instance
disk

2000 GB EBS GP2 IBM Spectrum 8

volume Protect database
disk

16384 GB EBS ST1 IBM Spectrum 2

volume Protect database
backup disk

665 GB EBS GP2 IBM Spectrum 1

volume Protect database
active log disk

4096 GB EBS ST1 IBM Spectrum 2

volume Protect database
archive log disk

3300 GB EBS GP2 IBM Spectrum 8

volume Protect cloud
cache disk

Object storage Amazon S3 bucket IBM Spectrum 1

Protect Amazon S3
bucket
(same AWS region
as instance)
Accessed via VPC
endpoint using
HTTPS

Table 4: AWS, medium configuration

Cloud component AWS component Detailed Quantity

description

Server and network m5.12xlarge or 48-core Intel 1

m5a.12xlarge Platinum 8175 @
Amazon EC2 3.10 GHz or AMD
instance (dedicated) EPYC 7000 @ 2.5
GHz

21
 192 GB RAM

EBS disk optimized

with 7000 Mbps
EBS throughput

10 Gbit Ethernet
connectivity

Virtual Private Instance placed in 1

Cloud (VPC) VPC; endpoint
created for Amazon
S3 service in the
same AWS region

Operating system Red Hat Enterprise 1

Linux or Ubuntu
Linux server

Block storage 100 GB EBS GP2 Operating system 1

volume disk

100 GB EBS GP2 IBM Spectrum 1

volume Protect instance
disk

1500 GB EBS GP2 IBM Spectrum 4

volume Protect database
disk

2048 GB EBS ST1 IBM Spectrum 2

volume Protect database
backup disk

512 GB EBS GP2 IBM Spectrum 1

volume Protect database
active log disk

1024 GB EBS ST1 IBM Spectrum 3

volume Protect database
archive log disk

3300 GB EBS GP2 IBM Spectrum 4

volume Protect cloud cache
disk
 Object storage Amazon S3 bucket IBM Spectrum 1
Protect Amazon S3
bucket
(same AWS Region
as instance)
Accessed via VPC
endpoint using
HTTPS

Table 5: AWS, small configuration

Cloud component AWS component Detailed Quantity

description

Server and network m5.4xlarge or 16-core Intel 1

m5a.4xlarge Platinum 8175 @
Amazon EC2 3.10GHz or AMD
instance (shared or EPYC 7000 @ 2.5
dedicated) GHz

64 GB RAM

EBS disk optimized

with 3500 Mbps
EBS throughput

Up to 10 Gbit
Ethernet
connectivity

Virtual Private Instance placed in 1

Cloud (VPC) VPC; endpoint
created for Amazon
S3 service in the
same AWS Region

Operating system Red Hat Enterprise 1

Linux or Ubuntu
Linux server

Block storage 100 GB EBS GP2 Operating system 1

volume disk

23
 100 GB EBS GP2 IBM Spectrum 1
volume Protect instance
disk

665 GB EBS GP2 IBM Spectrum 4

volume Protect database
disk

1024 GB EBS ST1 IBM Spectrum 2

volume Protect database
backup disk

512 GB EBS GP2 IBM Spectrum 1

volume Protect database
active log disk

1024 GB EBS ST1 IBM Spectrum 1

volume Protect database
archive log disk

3300 GB EBS GP2 IBM Spectrum 2

volume Protect cloud cache
disk

Object storage Amazon S3 bucket IBM Spectrum 1

Protect Amazon S3
bucket
(same AWS Region
as instance)
Accessed via VPC
endpoint using
HTTPS

Table 6: AWS, extra-small configuration

Cloud component AWS component Detailed Quantity

description

Server and network m5.xlarge or 4-core Intel 1

m5a.xlarge Amazon Platinum 8175 @
EC2 instance 3.10 GHz or AMD
(shared or EPYC 7000 @ 2.5
dedicated) GHz
 16 GB RAM

EBS disk optimized

with up to 2120
Mbps EBS
throughput

Up to 10 Gbit
Ethernet
connectivity

Virtual Private Instance placed in 1

Cloud (VPC) VPC; endpoint
created for Amazon
S3 service in the
same AWS Region

Operating system Red Hat Enterprise 1

Linux or Ubuntu
Linux server

Block storage 30 GB EBS GP2 Operating system 1

volume disk

30 GB EBS GP2 IBM Spectrum 1

volume Protect instance
disk

64 GB EBS GP2 IBM Spectrum 5 (5x file systems

volume Protect database created with Linux
and active log disk LVM)

512 GB EBS ST1 IBM Spectrum 1 (5x file systems

volume Protect database created with Linux
backup and archive LVM)
log disk

1200 GB EBS GP2 IBM Spectrum 1

volume Protect cloud cache
disk

Object storage Amazon S3 bucket IBM Spectrum 1

Protect Amazon S3
bucket
(same AWS Region
as instance)

25
 Accessed via VPC
endpoint using
HTTPS

2.1 Design Considerations for Amazon EC2 Instances

Several factors play into the choice of instance type and disk configuration for Amazon
instances. For each of the sizes, choose an instance type that will provide an acceptable
balance of vCPU and memory capacity to align with IBM Spectrum Protect sizing targets.
Regardless of size, any instance chosen should feature EBS optimizations with dedicated
bandwidth to the attached disk layer. This feature is deemed necessary, given the I/O
requirements of the IBM Spectrum Protect server database and the throughput and I/O
requirements of the cloud accelerator cache as the primary backup target. Each of the
preferred Amazon EC2 instances mentioned here are launched with EBS optimizations
enabled by default. For instances that do not feature EBS optimizations at launch by
default but that have this feature available, ensure that the EBS optimizations are enabled
when configuring the instance. See References [4].
Some Amazon EC2 instance lines tend to provide too much of one resource at excess cost
or too little of another resource than is needed by IBM Spectrum Protect. The other t2 and
m3 “General Purpose” instance lines were not sufficient from a vCPU and memory
perspective. The “Memory Optimized” x1 and r4 instances provide unnecessary amounts
of memory. The “Accelerated Computing” p3, p2, g3, and f1 instance lines, geared toward
machine learning and compute-intensive applications, are costly. Thus, they make a poor
choice for IBM Spectrum Protect deployments
Instance lines that could potentially match IBM Spectrum Protect requirements include the
i3 and d2 lines of “Storage Optimized” instances as well as the c4 line of “Compute
Optimized” instances and the m4 and m5 “General Purpose” instance lines.
Of the i3 line, the i3.4xlarge instance type is a candidate for small deployments; the
i3.8xlarge instance type is a candidate for medium deployments; and the i3.8xlarge and
i3.16xlarge instance types are candidates for large deployments. The i3.8xlarge instance
type is also capable enough from a vCPU and memory standpoint for a smaller-end large
configuration. At first glance, the i3 instance line’s internal non-volatile memory express
(NVMe) solid-state drive (SSD) storage could potentially be considered for use as cloud
accelerator cache storage. However, it should be cautioned that for all non-EBS volumes,
if you stop or terminate the instance, the disk volumes are erased. Even if you temporarily
stop the instance with the intention of starting it later, disk volumes are erased. When the
system restarts, its non-EBS volumes are provided anew, and any previous data that might
have resided on disks of this type are lost. Use of this potentially volatile storage should
be avoided as storage-pool directory targets due to the potential for data loss. Attached
disks (EBS) are the preferred and only reasonable repository for IBM Spectrum Protect
server data.
Of the d2 line, the d2.2xlarge instance type could serve for a small configuration, the
d2.4xlarge instance type for a medium configuration, and the d2.8xlarge instance type for a
large configuration. The d2 line features internal storage that does not involve the EBS
layer. The same caveats apply as mentioned for the i3 line.
Of the c4 line, the most appropriate instance type is the c4.8xlarge for small or a smaller-
end medium. None of the c4 line instance types satisfy the memory requirements of a large
system and should be avoided for that reason.
From a cost point of view, the m5.2xlarge instance type for small, the m5.4xlarge instance
type for medium, and the m5.10xlarge instance type for large are each much less
expensive than their i3, d2, or c4 counterparts in terms of hourly “On Demand” runtime
cost. However, of the m5 line, the m5.16xlarge instance type is best suited for a true large
system given its vCPU and memory capacity. For this reason, the m5.16xlarge instance
type is the current preferred choice.
In general, the current generation of the m5 line of Amazon EC2 instances is preferred
for large, medium, small, and extra-small IBM Spectrum Protect deployments. This newer
“general purpose” instance type provides a good combination of CPU, memory, Ethernet
network, and EBS disk performance with the latest Intel Platinum 8175 or AMD EPYC
7000 series processors. In addition, the m5 instances are typically more affordable than
their earlier generation m4 counterparts. Internal lab benchmarks have found that AMD
EPYC-based m5a instances provide better overall performance at scale with IBM
Spectrum Protect than Intel Platinum models, at a slightly lower cost.
If you choose an instance type that does not meet the memory requirements of a
Blueprint-level system, mitigation steps might be necessary to avoid out-of-memory
conditions when the system is under a Blueprint-level load. Mitigation steps can include
use of the DBMEMPERCENT server parameter to tune down maximum system Db2
database memory at the expense of database performance, reducing the maximum
number of parallel client sessions dispatched to the server, or both. You might apply
additional measures. If in doubt, adhere to the physical Blueprint targets.
Table 7: AWS instance options

Blue-
Sizing Instance Blueprint
Instance type vCPU Memory print
category line CPU
memory
Extra
m5(a) M5(a).xlarge - -
small 4 16 GB
m5(a) m5(a)xlarge 16 64 GB
Small i3 i3.4xlarge 16 122 GB 12 64 GB
d2 d2.2xlarge 8 61 GB
c4 c4.8xlarge 36 60 GB
m5(a) m5(a).12xlarge 48 192 GB
Medium 16 128 GB
i3 i3.8xlarge 32 244 GB
d2 d2.4xlarge 16 122 GB
Large m4 m4.10xlarge 40 160 GB 44

27
 i3 i3.8xlarge 32 244 GB
256 GB
m5(a) m5(a).16xlarge 64 256 GB (V3.1) or
i3 i3.16xlarge 64 488 GB 384 GB
(V4.1)
d2 d2.8xlarge 36 244 GB

The previous table is sorted from lowest to highest cost instance in each category. Costs
are estimated from the Amazon EC2 calculator based on a Red Hat Enterprise Linux
instance in the US West Oregon Region with On-Demand hourly pricing. Pricing is subject
to change; for this reason, exact figures are not provided. The entries highlighted in blue
are the current preferred Amazon EC2 instance types for extra-small, small, medium, and
large configurations, respectively. The blue entries provide a good balance between
controlling costs while satisfying IBM Spectrum Protect server requirements.
Beginning with IBM Spectrum Protect V8.1.3, the server automatically throttles client
backup operations if the cloud accelerator cache portion of a direct-to-cloud storage pool is
nearing full capacity. As a result, it is not mandatory to configure cloud accelerator disk
cache space that would be large enough to hold a full day’s worth of backups (after data
deduplication and compression). However, disk benchmarks should be run to ensure that
the anticipated back-end workload that an IBM Spectrum Protect server is expected to
support will not result in this disk location being the primary bottleneck of the system (see
Disk Benchmarking). In the previous table, the EC2 instances were carefully selected to
ensure that they have the dedicated EBS bandwidth and disk resources that are necessary
for the desired role. In practice, any planned deployment should be validated to ensure it
will meet performance requirements.
The NVMe SSD storage available for many instances, including the i3 line, should be
avoided for use by IBM Spectrum Protect. If the instance is stopped, all data in this non-
EBS disk will be lost. This would include a scenario in which data that is ingested by the
IBM Spectrum Protect server into the cloud accelerator cache, and reported successfully
stored to the client, is lost from this disk before the asynchronous disk-to-cloud transfer is
completed.
Thresholds, particularly at the EBS disk level, must be kept in mind when designing and
deploying complete instance solutions. For example, IOPS are limited to a count of 80,000
per AWS account per disk type. Failure to account for IOPS or throughput limitations
imposed by the service provider on the account, instance, or disk level can result in
performance problems that might be difficult to debug in production. For Amazon
documentation about this subject, see References [4] [5].

2.1.1 Considerations for Direct-to-Cloud Architectures

The direct-to-cloud configurations discussed here are architected such that the disk of the
cloud accelerator cache is on fast-performing SSD. In general, the limiting factor for end-to-
end ingestion throughput for an IBM Spectrum Protect server using object storage is the
network to the object storage system or the object storage system itself. For this reason,
you must understand the throughput capability requirements of the planned system. Those
 requirements will drive your decisions about assigning servers to Ethernet networks. Next,
the ceiling for daily ingestion throughput (in terms of mebibytes per second or tebibytes per
day) must be determined at the object storage end by using object storage
benchmarking tests. The test results will establish what the overlapped I/O capability
would have to be for the cloud accelerator cache disk location for maximum back-end
throughput. These two factors will help you select a disk technology to sustain daily backup
ingestion requirements while working within object storage limits. After you select a disk
technology, run disk benchmarking tests to ensure throughput capability (see Disk
Benchmarking).
Tip: In this paper, the abbreviation MiB is used for mebibytes, the abbreviation TiB is used
for tebibytes, the abbreviation KiB is used for kibibytes, and the abbreviation GiB is used
for gibibytes.
Example
If an object storage link is capable of 10 Gbps, this data transfer speed equals about 1000
MiB/s after packet overhead and other efficiency loss. In order to saturate this link for long
periods, the cloud accelerator cache disk location must be capable of taking in client
ingestion data (writes) at 1000 MiB/s and transmitting staged data to object storage (reads)
at a similar speed, 1000 MiB/s (~128-256 KiB I/O size). This capacity ensures that the
cloud accelerator cache disk can remain as small as possible while sustaining maximum
throughput. Alternatively, a larger capacity, slower disk technology (such as Amazon EBS
st1 magnetic disk) can be used such that the client ingestion data that has been staged to
accelerator disk cache can be transmitted to object storage over a longer period of the day
(extending past the backup window). However, beware that data residing only in the cloud
accelerator cache is unprotected in the sense that only a single copy of the data exists.
The redundancy protection inherent in cloud object storage is available only if the data is
transmitted to object storage. Generally, Amazon EBS disks provide acceptable durability.

2.1.2 Sizing the Cloud Accelerator Cache

Figure 4 can be used as a rough guide for the appropriate disk technology to use based on
object storage and object storage network capability. At the top left, Amazon S3 object
storage is reachable over the same LAN (for example, within the same AWS Virtual Private
Cloud network). As we move from top to bottom in the diagram, the network capability
becomes slower (10 Gbps to 1 Gbps), while the storage capacity requirements increase to
store data that is queued up in the accelerator disk cache awaiting transfer to object
storage. In the case of slower network-to-object storage, it is more likely that
(asynchronous) client ingestion from local client systems can run at a faster rate than cloud
transfer. In such a scenario, client ingestion data begins to fill the cache disk location faster
than the data can be transferred to object storage and cleared from the cache. As of IBM
Spectrum Protect V8.1.2, an internal throttling mechanism is in place to slow client
ingestion speeds if the cloud accelerator cache disk area begins nearing capacity.
However, to avoid slowing client ingestion in cases where ingestion exceeds the cloud
transfer rate (which might not be desired), the accelerator cache should be sized with a
larger capacity, perhaps up to a single day’s worth of back-end client ingestion (after data
deduplication and compression).

29
 Figure 4: Sizing the cloud accelerator cache for AWS

2.1.3 AWS: Large Instance Considerations

For a large Amazon based build, the m5(a).16xlarge instance type is the preferred choice.
The newer generation m5 or m5a “general purpose” instance lines feature updated Intel
Platinum or AMD EPYC processors on the AWS Nitro System hypervisor and are generally
more affordable than their earlier m4 counterparts. The AWS Nitro System instances are
generally higher-performing instance types with lines including m5, c5, and r5 instance
types. For more information, see References [9].
Amazon GP2 type EBS disks were chosen for the IBM Spectrum Protect database disk
and active log disk volumes as a more cost-effective alternative to io1 type disks. To meet
IOPS requirements of the Db2 component, a larger capacity disk is required to enable the
increased IOPS capability of these underlying “general purpose” EBS disks. Eight disks at
2000 GB each for the database volumes and one disk at 665 GB for the active log were
sufficient for this purpose. A capacity of approximately 15.6 TB represents an
overprovisioning of space for the database in exchange for the required IOPS scalability.
Amazon st1 type throughput-optimized EBS disks were chosen for use as IBM Spectrum
Protect database backup and database archive log volumes due to their sequentially
optimized nature and cost effectiveness.
Amazon gp2 type EBS disks were chosen for cloud accelerator cache due to their
efficient random IOPS characteristics. Configurations with these types of disks seem to be
ideal for handling the direct-to-cloud pattern of 256 KB sequential writes mixed with
asynchronous 128 KB, somewhat sequential reads (of other files on disk). Eight of these
disks at a 3300 GB capacity each were chosen to provide enough overlapped throughput
capability to allow the IBM Spectrum Protect server to saturate the back-end link to S3
storage from this instance. A total capacity of more than 25 TB results from this
configuration, providing enough storage for a single day’s worth of large Blueprint-level
 backup ingestion, assuming a favorable (for example, 4:1) data reduction ratio with
deduplication and compression of front-end client data.
To review disk benchmarks for database disk volumes and cloud accelerator cache
volumes for all built configurations, see Disk Benchmarking.

2.1.4 AWS: Medium Instance Considerations

For the medium Amazon build, the m5(a).12xlarge instance types are preferred (Intel or
AMD variants). This instance type provides 48 vCPUs and 192 GB of RAM with 875 MB/s
of advertised dedicated EBS throughput and 10 Gbit Ethernet network capability. This
instance provides for more CPU and memory capacity than would typically be required of a
medium Blueprint-level system, allowing this instance type to be an option for a smaller-
end large build as well. Other alternatives for the medium Blueprint role tend to provide
less memory than is preferred, and this factor could lead to resource constraints on a
loaded system in production. More cost-effective alternatives to the medium role could be
pursued within EC2, with the caveat that memory demands and usage of the IBM
Spectrum Protect server and database should be closely monitored during high load.
As with the large instance type, Amazon gp2 disks were again chosen for the primary
database volumes and active log volume, but at a smaller capacity for both. A total
database size of approximately 5.8 TB is provided with this configuration. Once again,
excess capacity is provisioned with these less expensive disks to achieve the desired
IOPS capability. st1 disks were again chosen for the server’s database backup and archive
log disk volumes.
For the cloud accelerator cache disk, gp2 volumes were again chosen. However, only four
are needed for a medium configuration.

2.1.5 AWS: Small Instance Considerations

For the small Amazon build, the m5(a).4xlarge instance types are preferred (Intel or AMD
variants). This instance type provides 16 vCPUs and 122 GB of RAM, providing more
server memory than the physical Blueprint specification. This instance type provides for
437.5 MB/s of dedicated throughput to the EBS disk layer, along with up to 10 Gbit
Ethernet bandwidth capability. A similar caveat applies with this instance as with the
medium EC2 instance: options are present that could serve the role of an IBM Spectrum
Protect medium deployment but with less memory capability than is preferred, at the risk of
being memory constrained on the system during periods of high load.
Again, gp2 type EBS disks were used here for database and active log volumes, providing
2.6 TB of database capacity across four database disks. Disk capacities for the st1-based
database backup and archive log volumes were also reduced.
In this case, only two 3300 GB gp2 type volumes were used for cloud accelerator cache,
providing a total capacity of 6600 GB.

2.1.6 AWS: Extra-Small Instance Considerations

For the extra-small Amazon build, the m5(a).xlarge instance types are preferred (Intel or
AMD variants). This instance type features 4 vCPUs and 64 GB of RAM. For maximum
cost savings, the m5a.xlarge AMD variant and Ubuntu Linux could be chosen. This

31
 instance type provides up to 2120 Mbps of dedicated throughput to the EBS disk layer,
along with up to 10 Gbit Ethernet bandwidth.
EBS gp2 disks were again chosen for the database and active log volumes to ensure
adequate IOPs capability, even for this smaller instance size. For convenience with
scalability, the five 64 GB disks were combined into a single LVM volume group with four
database volumes and one active log volume created. Similarly, a single 512 GB ST1
volume was provisioned and divided into two volumes using LVM, one for database
backup and one for use as the archive log volume. If additional database backup capacity
is required, additional EBS disk capacity could be added to this volume group. This applies
as well to the one 1200 GB GP2 disk provisioned for the cloud accelerator cache. General
purpose SSD performance is still preferable for the overlapped I/O seen with this volume.

Throughput Measurements and Results

Throughput measurements for backup and restore operations are provided with
representative datasets for selected configurations. All throughput measurements were
collected with IBM Spectrum Protect servers configured in a direct-to-cloud, accelerator
cache-backed setup, with primary backup data stored to object storage. Throughput
measurements with a disk-to-cloud tiering configuration are not included in this document.
For each configuration tested, the following preferred settings were adopted for the cloud-
container storage pool. To optimize performance when you deploy an IBM Spectrum
Protect server instance in the cloud, use the following settings:

• Storage pool compression enabled

• Storage pool encryption enabled

• Storage pool configured as an off-premises cloud-container storage pool

• HTTPS URLs specified for all cloud endpoints

3.1 Dataset Descriptions

For the performance tests that were conducted within these environments, all or a subset
of the following datasets were used:
Table 8: Throughput measurement datasets

Front-end Average Duplicate data Data Notes

object size duplication percentage compressibility
extent size

128 MiByte ~150-200 ~70% ~50% VE-like, favorable

KiByte extent size

128 MiByte ~200-300 ~0% ~0% Random data,

KiByte large extent size
 1 GiByte ~60-100 ~50% ~50% DB-like, small
KiByte extent size

128 KiByte ~128 KiByte ~0% ~95% Small file, high

overhead

The 128 MiByte, VE-like front-end dataset represents a relatively large object size that
aligns with the IBM Spectrum Protect for Virtual Environments: Data Protection for VMware
API client’s VE megablock size for virtual machine disk backups. The large object size
and relatively large, though realistic, deduplication extent size represents a favorable
profile for the IBM Spectrum Protect server’s ingestion engine to achieve good
performance. A duplicate data rate of 70% combined with a compressibility rate of 50% for
this dataset yields an 85% total data reduction from front-end data as compared with data
that is actually stored to the (cloud-accelerator cache and object storage) back-end after
data deduplication, compression, and encryption processing. Although this workload does
not qualify as a “best case,” it does represent a realistic, favorable scenario in which to
model top-end throughput capability of an IBM Spectrum Protect system without
overestimating throughput performance.
The 128 MiByte, random front-end dataset represents a larger object size with a large,
favorable deduplication extent size. However, the random nature of the data ensures that it
does not deduplicate well with existing storage pool data or compress well. This dataset is
included to represent a workload that is throughput intensive from the perspective of
storage pool disk and object storage network load. Full backups of large objects containing
relatively random data content would be modeled well by this dataset.
The 1 GiByte front-end dataset represents a model of structured, database-like data
possessing a relatively small deduplication extent size relative to the front-end object size.
Such a workload is representative of what might be experienced with an IBM Spectrum
Protect for Databases: Data Protection for Oracle backup environment protecting
production databases. The smaller extent size causes additional strain and overhead for
the IBM Spectrum Protect ingestion engine and typically results in less throughput than the
128 MiByte dataset. A duplicate data rate of 50% and compressibility of 50% yield a 75%
overall front-end to back-end reduction for this workload, with a 4:1 ratio reduction, which
approaches what is seen for this type of data in the field.
The 128 KiByte front-end dataset is used here as a relative “worst case” workload in
regard to throughput efficiency for the IBM Spectrum Protect server. This “small file”
workload will stress the data deduplication, compression, encryption, and object storage
transfer components of the IBM Spectrum Protect server to a higher degree relative to the
amount of data that is actually protected and transferred to object storage. This high
overhead dataset allows for predicting a lower estimate on performance of an IBM
Spectrum Protect environment within these cloud solutions.

33
3.2 Backup and Restore Measurements
The following sections outline the backup and restore throughput results that were
experienced with the previously mentioned datasets in the built cloud environments.
Prior to conducting backup and restore tests on the IBM Spectrum protect environments, a
load phase was conducted whereby the servers were initially loaded with a set of
deduplicated 128 MiByte front-end data in order to populate the server database tables to
provide for a more realistic customer configuration. IBM Spectrum Protect database
queries can change their behavior based on the size and layout of server database tables.
This load phase was performed to bring behavior in line with real environment
expectations.
For each dataset, up to 50 IBM Spectrum Protect client backup sessions were initiated in
parallel to the server for the large Amazon configuration, up to 25 for the medium Amazon
configuration, and up to 10 for the small Amazon configuration. The results presented here
for backup represent the maximum front-end throughput experienced with the largest
number of sessions tested against that system.
For each dataset on restore, between 1 and 40 client restore sessions were initiated for the
large Amazon system, between 1 and 20 for the medium Amazon system, and between 1
and 10 for the small Amazon system. Results presented here include the intermediate
session count values to show how restore throughput can scale with the number of restore
sessions involved for datasets similar to these types.
All throughput values represent front-end, “protected data” values, before inline data
deduplication, compression, and encryption. These are the data rates experienced by a
client that is backing up data to or restoring data from the IBM Spectrum Protect server.
The rates are similar to what customers would likely describe as their performance
experience with the product. On ingestion, the actual quantity of data that makes it to
accelerator cache disk and onwards to object storage will be less, depending on the data
deduplication and compression rate. On restore, all individual extents comprising a front-
end object will be restored using HTTP GET calls from the object storage device. However,
the built-in caching within the IBM Spectrum Protect server’s restore engine might reduce
the number of restore operations needed if a workload contains duplicate data.

3.2.1 AWS: Large Instance Measurements

128 MiByte VE-like front-end dataset results:

Table 9: AWS, large configuration, 128 MiByte VE-like dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

50 70% 2147.8 7.4 59.0 73.7

For the large Amazon build, front-end aggregate throughput for the 50-session favorable
dataset achieved 6.9 TiBytes/hour, which would yield capability well within the large
Blueprint range. Values are provided here for estimates using both 8-hour and 10-hour
backup windows in a given day. Both estimates are provided for convenience to align with
our experience that customers generally use one of these two time periods for their daily
ingestion workloads.

Table 10: AWS, large configuration, 128 MiByte VE-like dataset restore

Sessions MiBytes/s Gibytes/hour

1 209.7 737.1

2 392.3 1379.0

4 627.7 2206.9

8 1107.1 3892.3

16 1128.4 3966.9

32 1736.8 6105.9

With the 50% compressibility of this dataset, the amount of data transmitted over the wire
(from object storage) is half its decompressed front-end value. This allows for throughput in
excess of 10 Gbit line speed without saturating the EC2-to-S3 Ethernet link. Restore rates
from object storage are typically challenged due to latency and individual read
characteristics. Restore throughput rates with realistic session counts from object storage
pools might challenge certain data recovery time objectives (RTOs).

128 MiByte random front-end dataset results:

Table 11: AWS, large configuration, 128 MiByte random dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

50 0% 296.2 1.0 8.1 10.2

Table 12: AWS, large configuration, 128 MiByte random dataset restore

35
 Sessions MiBytes/s Gibytes/hour

1 145.0 509.8

2 253.8 892.4

4 385.4 1354.9

8 477.9 1680.0

16 540.2 1899.1

32 405.2 1424.7

This dataset benefits from a large deduplication extent size of approximately 200 - 300
KiBytes, on average. This allows for more data movement per HTTP GET operation
performed by the cloud-container storage pool and subsequently greater throughput per
session compared to the previous 128 MiByte dataset.

1 GiByte front-end dataset results:

Table 13: AWS, large configuration, 1 GiByte dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

50 50% 1538.1 5.3 42.2 52.8

Ingestion throughput rates with the 1 GiByte dataset featuring the smaller (60 – 100
KiByte) average deduplication extent size were less than the 128 MiByte favorable
workload, but still in line with large Blueprint throughput targets in this direct-to-cloud
configuration.

Table 14: AWS, large configuration, 1 GiByte dataset restore results

Sessions MiBytes/s Gibytes/hour

1 129.9 456.7

2 235.3 827.4

4 291.4 1024.3
 8 446.0 1568.1

16 346.9 1219.5

32 477.4 1678.5

Restore throughput rates with the smaller extent workload are somewhat less than the
favorable, 128 MiByte dataset. This is due to the additional overhead of more HTTP GET
requests necessary to restore this data (smaller extents lead to more operations per front-
end restored data). A throughput of more than 1 TiByte/hour could be achieved with 20-40
restore sessions of this data type.

3.2.2 AWS: Medium Instance Measurements

128 MiByte VE-like front-end dataset results:
Table 15: AWS, medium configuration, 128 MiByte VE-like dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

25 70% 1108.9 3.8 30.5 38.1

Throughput with the favorable 128 MiByte object dataset with the medium Amazon build
easily falls within the upper end of the medium Blueprint throughput target with 25 backup
sessions for an 8-hour backup window and surpasses it with a 10-hour window. This
configuration could possibly support the data workload of a smaller, large Blueprint role.

Table 16: AWS, medium configuration, 128 MiByte VE-like dataset restore results

Sessions MiBytes/s GiBytes/hour

1 254.4 894.5

2 506.4 1780.4

4 859.3 3020.9

8 1823.2 6409.6

16 1910.4 6716.2

37
128 MiByte random front-end dataset results:
Table 17: AWS, medium configuration, 128 MiByte random dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

25 0% 207.0 0.7 5.7 7.1

Table 18: AWS, medium configuration, 128 MiByte random dataset restore results

Sessions MiBytes/s GiBytes/hour

1 163.7 575.4

2 268.3 943.2

4 419.1 1473.5

8 444.1 1561.2

16 463.8 1630.5

1 GiByte front-end dataset results:

Table 19: AWS, medium configuration, 1 GiByte dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

25 50% 775.8 2.7 21.3 26.6

With the 1 GiByte dataset workload, ingestion throughput for 8- or 10-hour backup
windows is still within medium Blueprint targets (6 – 20 TiBytes per day front-end data
protected).

Table 20: AWS, medium configuration, 1 GiByte dataset restore results

Sessions MiBytes/s GiBytes/hour

1 157.4 553.2
 2 313.1 1100.7

4 644.4 2265.5

8 593.0 2084.7

16 578.8 2034.8

3.2.3 AWS: Small Instance Measurements

128 MiByte VE-like front-end dataset results:
Table 21: AWS, small configuration, 128 MiByte VE-like dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

10 70% 825.8 2.8 22.7 28.4

Table 22: AWS, small configuration, 128 MiByte VE-like dataset restore results

Sessions MiBytes/s GiBytes/hour

1 226.2 795.1

2 415.2 1459.6

4 644.1 2264.5

8 887.9 3121.6

128 MiByte random front-end dataset results:

Table 23: AWS, small configuration, 128 MiByte random dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

10 0% 104.1 0.4 2.9 3.6

39
 Table 24: AWS, small configuration, 128 MiByte random dataset restore results

Sessions MiBytes/s GiBytes/hour

1 141.0 495.7

2 223.4 785.4

4 286.1 1005.9

8 339.5 1193.7

1 GiByte front-end dataset results:

Table 25: AWS, small configuration, 1 GiByte dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

10 50% 808.6 2.8 22.2 27.8

Table 26: AWS, small configuration, 1 GiByte dataset restore results

Sessions MiBytes/s GiBytes/hour

1 140.1 492.7

2 244.1 858.1

4 336.0 1181.3

8 391.3 1375.6

3.2.4 AWS: Extra-Small Instance Measurements

128 MiByte VE-like front-end dataset results:
Table 27: AWS, extra-small configuration, 128 MiByte VE-like dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

5 70% 306.9 1.1 8.4 10.5

Table 28: AWS, extra-small configuration, 128 MiByte VE-like dataset restore results

Sessions MiBytes/s GiBytes/hour

1 180.1 633.2

2 330.7 1162.6

4 555.4 1952.6

128 MiByte random front-end dataset results:

Table 29: AWS, extra-small configuration, 128 MiByte random dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

5 0% 50.0 0.2 1.4 1.7

Table 30: AWS, extra-small configuration, 128 MiByte random dataset restore results

Sessions MiBytes/s GiBytes/hour

1 96.7 339.8

2 109.5 384.8

4 122.1 429.1

1 GiByte front-end dataset results:

Table 31: AWS, extra-small configuration, 1 GiByte dataset backup results

Sessions Data MiBytes/s TiBytes/hour TiBytes per TiBytes per

deduplication 8 hours 10 hours
%

5 50% 267.1 0.9 7.3 9.2

Table 32: AWS, extra-small configuration, 1 GiByte dataset restore results

41
 Sessions MiBytes/s GiBytes/hour

1 120.4 423.4

2 172.2 605.4

4 190.6 670.1

APPENDIX

Disk Setup Commands for Linux

Deployments
For IBM Spectrum Protect deployments on AWS cloud computing systems, the preferred
operating system is Linux, either the latest IBM Spectrum Protect supported Red Hat
Enterprise Linux or Ubuntu Linux. With Ubuntu Linux, care should be taken to ensure that
all required Linux packages are installed to enable Linux Logical Volume Manager (LVM)
functionality. For more information about the operating systems, see the IBM Spectrum
Protect technote (References [7]).
This chapter provides reference information for deployments of IBM Spectrum Protect with
Ubuntu Linux 16.04 LTS for large, medium, small, and extra-small environments. In this
case, Ubuntu Linux was installed on each instance on the Amazon EC2 cloud computing
system. The following sections outline the disk, file system, and miscellaneous commands
that were executed after launching the instance in preparation for installing IBM Spectrum
Protect. These sections reflect the preferred instance type and direct-to-cloud disk
configurations that are outlined in Amazon EC2 Compute Configurations.
If you use Linux LVM and logical volumes for the cloud accelerator cache (and sometimes
archive log) disk, ensure that LVM striping is used with the lvcreate command such
that logical volumes make use of data stripes from all physical disks in the volume group.
The number of stripes that you specify when creating a logical volume should match the
number of physical disks in the volume group. Lab testing indicates that a suitable stripe
size for cloud accelerator cache and archive log purposes is 16 KiBytes. For example, for
a cloud accelerator cache volume group making use of two physical disks, you could use
the following command:
lvcreate --stripes 2 --stripesize 16 --extents 100%FREE --name sp_cc_vg sp_cc

IBM Spectrum Protect, Large Amazon EC2 System, Ubuntu Linux: m5a.16xlarge
Instance Type

#############################################

# Miscellaneous

#############################################

# Update apt packages, install tooling

apt update

apt upgrade -y

apt install scsitools -y

apt install dstat -y

apt install sysstat -y

apt install multipath-tools -y

apt install multipath-tools-boot -y

# Start the multipath daemon

touch /etc/multipath.conf

systemctl restart multipath-tools.service

# Set server hostname

hostnamectl set-hostname <desired hostname>

# Set vm.swappiness to 5 to match v4.1 xLinux blueprint recommendation

sysctl vm.swappiness=5

43
# Install the Korn shell

apt install ksh -y

# Reboot the system

#############################################

# Spectrum Protect Disk Setup (Ubuntu)

#############################################

# Run the following as the "root" user

# Install LVM for Ubuntu

apt install lvm2 -y

systemctl enable lvm2-lvmetad

systemctl start lvm2-lvmetad

# Make the required Spectrum Protect directories

mkdir /sp

mkdir /sp/tsminst1

mkdir /sp/sp_db1

mkdir /sp/sp_db2

mkdir /sp/sp_db3

mkdir /sp/sp_db4

mkdir /sp/sp_db5

mkdir /sp/sp_db6

mkdir /sp/sp_db7

mkdir /sp/sp_db8

mkdir /sp/sp_dbb1

mkdir /sp/sp_dbb2

mkdir /sp/sp_alog

mkdir /sp/sp_archlog

mkdir /sp/sp_cc

# Note that disk name values may be different from those seen here in actual
deployments

# Instance directory

# Change the value for "nvme*" to match your environment

mkfs.xfs /dev/nvme8n1

mount /dev/nvme8n1 /sp/tsminst1

# Cloud Cache

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme4n1

pvcreate /dev/nvme6n1

pvcreate /dev/nvme7n1

pvcreate /dev/nvme10n1

pvcreate /dev/nvme12n1

pvcreate /dev/nvme14n1

pvcreate /dev/nvme18n1

pvcreate /dev/nvme19n1

vgcreate sp_cc /dev/nvme4n1 /dev/nvme6n1 /dev/nvme7n1 /dev/nvme10n1 /dev/nvme12n1

/dev/nvme14n1 /dev/nvme18n1 /dev/nvme19n1

lvcreate --stripes 8 --stripesize 16 --extents 100%FREE --name sp_cc_vg sp_cc

mkfs.xfs /dev/mapper/sp_cc-sp_cc_vg

mount /dev/mapper/sp_cc-sp_cc_vg /sp/sp_cc

# DB

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme1n1

pvcreate /dev/nvme3n1

pvcreate /dev/nvme5n1

pvcreate /dev/nvme9n1

pvcreate /dev/nvme11n1

pvcreate /dev/nvme15n1

pvcreate /dev/nvme16n1

pvcreate /dev/nvme22n1

vgcreate sp_db /dev/nvme1n1 /dev/nvme3n1 /dev/nvme5n1 /dev/nvme9n1 /dev/nvme11n1

/dev/nvme15n1 /dev/nvme16n1 /dev/nvme22n1

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db1 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db2 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db3 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db4 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db5 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db6 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db7 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 511999 --name sp_db_db8 sp_db

mkfs.ext4 /dev/mapper/sp_db-sp_db_db1

mkfs.ext4 /dev/mapper/sp_db-sp_db_db2

45
mkfs.ext4 /dev/mapper/sp_db-sp_db_db3

mkfs.ext4 /dev/mapper/sp_db-sp_db_db4

mkfs.ext4 /dev/mapper/sp_db-sp_db_db5

mkfs.ext4 /dev/mapper/sp_db-sp_db_db6

mkfs.ext4 /dev/mapper/sp_db-sp_db_db7

mkfs.ext4 /dev/mapper/sp_db-sp_db_db8

mount /dev/mapper/sp_db-sp_db_db1 /sp/sp_db1

mount /dev/mapper/sp_db-sp_db_db2 /sp/sp_db2

mount /dev/mapper/sp_db-sp_db_db3 /sp/sp_db3

mount /dev/mapper/sp_db-sp_db_db4 /sp/sp_db4

mount /dev/mapper/sp_db-sp_db_db5 /sp/sp_db5

mount /dev/mapper/sp_db-sp_db_db6 /sp/sp_db6

mount /dev/mapper/sp_db-sp_db_db7 /sp/sp_db7

mount /dev/mapper/sp_db-sp_db_db8 /sp/sp_db8

# Active Log

# Change the value for "nvme*" to match your environment

mkfs.ext4 /dev/nvme21n1

mount /dev/nvme21n1 /sp/sp_alog

# DB backup

# Change the value for "nvme*" to match your environment

mkfs.ext4 /dev/nvme13n1

mkfs.ext4 /dev/nvme20n1

mount /dev/nvme13n1 /sp/sp_dbb1

mount /dev/nvme20n1 /sp/sp_dbb2

# Archive log

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme2n1

pvcreate /dev/nvme17n1

vgcreate sp_archlog /dev/nvme2n1 /dev/nvme17n1

lvcreate --stripes 2 --stripesize 16 --extents 100%FREE --name sp_archlog_vg

sp_archlog

mkfs.ext4 /dev/mapper/sp_archlog-sp_archlog_vg

mount /dev/mapper/sp_archlog-sp_archlog_vg /sp/sp_archlog

# Edit /etc/fstab to ensure that mount points will be restored on reboot

vi /etc/fstab
#############################################

# Spectrum Protect Instance User Setup

#############################################

# Run the following as the "root" user.

# Add a Linux group and user to own the Spectrum Protect instance.

# Set the password for this instance.

groupadd tsmsrvrs

useradd -g tsmsrvrs tsminst1

passwd tsminst1

# Ensure that this user can use the mount points previously created

chown -R tsminst1:tsmsrvrs /sp/

# From here, proceed with the IBM Spectrum Protect installation and setup

IBM Spectrum Protect, Medium Amazon EC2 System, Ubuntu Linux: m5a.12xlarge
Instance Type

#############################################

# Miscellaneous

#############################################

# Update apt packages, install tooling

apt update

apt upgrade -y

apt install scsitools -y

apt install dstat -y

apt install sysstat -y

apt install multipath-tools -y

apt install multipath-tools-boot -y

# Start the multipath daemon

touch /etc/multipath.conf

systemctl restart multipath-tools.service

47
# Set server hostname

hostnamectl set-hostname <desired hostname>

# Set vm.swappiness to 5 to match v4.1 xLinux blueprint recommendation

sysctl vm.swappiness=5

# Install the Korn shell

apt install ksh -y

# Reboot the system

#############################################

# Spectrum Protect Disk Setup (Ubuntu)

#############################################

# Run the following as the "root" user

# Install LVM for Ubuntu

apt install lvm2 -y

systemctl enable lvm2-lvmetad

systemctl start lvm2-lvmetad

# Make the required Spectrum Protect directories

mkdir /sp

mkdir /sp/tsminst1

mkdir /sp/sp_db1

mkdir /sp/sp_db2

mkdir /sp/sp_db3

mkdir /sp/sp_db4

mkdir /sp/sp_dbb1

mkdir /sp/sp_dbb2

mkdir /sp/sp_alog

mkdir /sp/sp_archlog

mkdir /sp/sp_cc

# Note that disk name values may be different from those seen here in actual
deployments

# Instance directory

# Change the value for "nvme*" to match your environment

mkfs.xfs /dev/nvme6n1

mount /dev/nvme6n1 /sp/tsminst1

# Cloud Cache

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme2n1

pvcreate /dev/nvme5n1

pvcreate /dev/nvme9n1

pvcreate /dev/nvme14n1

vgcreate sp_cc /dev/nvme2n1 /dev/nvme5n1 /dev/nvme9n1 /dev/nvme14n1

lvcreate --stripes 4 --stripesize 16 --extents 100%FREE --name sp_cc_vg sp_cc

mkfs.xfs /dev/mapper/sp_cc-sp_cc_vg

mount /dev/mapper/sp_cc-sp_cc_vg /sp/sp_cc

# DB

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme8n1

pvcreate /dev/nvme11n1

pvcreate /dev/nvme12n1

pvcreate /dev/nvme15n1

vgcreate sp_db /dev/nvme8n1 /dev/nvme11n1 /dev/nvme12n1 /dev/nvme15n1

lvcreate --stripes 4 --stripesize 16 --extents 383999 --name sp_db_db1 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 383999 --name sp_db_db2 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 383999 --name sp_db_db3 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 383996 --name sp_db_db4 sp_db

mkfs.ext4 /dev/mapper/sp_db-sp_db_db1

mkfs.ext4 /dev/mapper/sp_db-sp_db_db2

mkfs.ext4 /dev/mapper/sp_db-sp_db_db3

mkfs.ext4 /dev/mapper/sp_db-sp_db_db4

mount /dev/mapper/sp_db-sp_db_db1 /sp/sp_db1

mount /dev/mapper/sp_db-sp_db_db2 /sp/sp_db2

mount /dev/mapper/sp_db-sp_db_db3 /sp/sp_db3

mount /dev/mapper/sp_db-sp_db_db4 /sp/sp_db4

# Active Log

# Change the value for "nvme*" to match your environment

mkfs.ext4 /dev/nvme3n1

mount /dev/nvme3n1 /sp/sp_alog

49
# DB backup

# Change the value for "nvme*" to match your environment

mkfs.ext4 /dev/nvme4n1

mkfs.ext4 /dev/nvme13n1

mount /dev/nvme4n1 /sp/sp_dbb1

mount /dev/nvme13n1 /sp/sp_dbb2

# Archive log

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme1n1

pvcreate /dev/nvme7n1

pvcreate /dev/nvme10n1

vgcreate sp_archlog /dev/nvme1n1 /dev/nvme7n1 /dev/nvme10n1

lvcreate --stripes 3 --stripesize 16 --extents 100%FREE --name sp_archlog_vg

sp_archlog

mkfs.ext4 /dev/mapper/sp_archlog-sp_archlog_vg

mount /dev/mapper/sp_archlog-sp_archlog_vg /sp/sp_archlog

# Edit /etc/fstab to ensure that mount points will be restored on reboot

vi /etc/fstab

#############################################

# Spectrum Protect Instance User Setup

#############################################

# Run the following as the "root" user.

# Add a Linux group and user to own the Spectrum Protect instance.

# Set the password for this instance.

groupadd tsmsrvrs

useradd -g tsmsrvrs tsminst1

passwd tsminst1

# Ensure that this user can use the mount points previously created

chown -R tsminst1:tsmsrvrs /sp/

# From here, proceed with the IBM Spectrum Protect installation and setup
IBM Spectrum Protect, Small Amazon EC2 System, Ubuntu Linux: m5a.4xlarge
Instance Type
#############################################

# Miscellaneous

#############################################

# Update apt packages, install tooling

apt update

apt upgrade -y

apt install scsitools -y

apt install dstat -y

apt install sysstat -y

apt install multipath-tools -y

apt install multipath-tools-boot -y

# Start the multipath daemon

touch /etc/multipath.conf

systemctl restart multipath-tools.service

# Set server hostname

hostnamectl set-hostname <desired hostname>

# Set vm.swappiness to 5 to match v4.1 xLinux blueprint recommendation

sysctl vm.swappiness=5

# Install the Korn shell

apt install ksh -y

# Reboot the system

#############################################

# Spectrum Protect Disk Setup (Ubuntu)

#############################################

# Run the following as the "root" user

# Install LVM for Ubuntu

apt install lvm2 -y

systemctl enable lvm2-lvmetad

51
systemctl start lvm2-lvmetad

# Make the required Spectrum Protect directories

mkdir /sp

mkdir /sp/tsminst1

mkdir /sp/sp_db1

mkdir /sp/sp_db2

mkdir /sp/sp_db3

mkdir /sp/sp_db4

mkdir /sp/sp_dbb1

mkdir /sp/sp_dbb2

mkdir /sp/sp_alog

mkdir /sp/sp_archlog

mkdir /sp/sp_cc

# Note that disk name values may be different from those seen here in actual
deployments

# Instance directory

# Change the value for "nvme*" to match your environment

mkfs.xfs /dev/nvme6n1

mount /dev/nvme6n1 /sp/tsminst1

# Cloud Cache

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme1n1

pvcreate /dev/nvme10n1

vgcreate sp_cc /dev/nvme1n1 /dev/nvme10n1

lvcreate --stripes 2 --stripesize 16 --extents 100%FREE --name sp_cc_vg sp_cc

mkfs.xfs /dev/mapper/sp_cc-sp_cc_vg

mount /dev/mapper/sp_cc-sp_cc_vg /sp/sp_cc

# DB

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme3n1

pvcreate /dev/nvme4n1

pvcreate /dev/nvme8n1

pvcreate /dev/nvme11n1

vgcreate sp_db /dev/nvme3n1 /dev/nvme4n1 /dev/nvme8n1 /dev/nvme11n1

lvcreate --stripes 4 --stripesize 16 --extents 170239 --name sp_db_db1 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 170239 --name sp_db_db2 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 170239 --name sp_db_db3 sp_db

lvcreate --stripes 4 --stripesize 16 --extents 170236 --name sp_db_db4 sp_db

mkfs.ext4 /dev/mapper/sp_db-sp_db_db1

mkfs.ext4 /dev/mapper/sp_db-sp_db_db2

mkfs.ext4 /dev/mapper/sp_db-sp_db_db3

mkfs.ext4 /dev/mapper/sp_db-sp_db_db4

mount /dev/mapper/sp_db-sp_db_db1 /sp/sp_db1

mount /dev/mapper/sp_db-sp_db_db2 /sp/sp_db2

mount /dev/mapper/sp_db-sp_db_db3 /sp/sp_db3

mount /dev/mapper/sp_db-sp_db_db4 /sp/sp_db4

# Active Log

# Change the value for "nvme*" to match your environment

mkfs.ext4 /dev/nvme9n1

mount /dev/nvme9n1 /sp/sp_alog

# DB backup and Archive log

# Change the value for "nvme*" to match your environment

mkfs.ext4 /dev/nvme2n1

mkfs.ext4 /dev/nvme5n1

mkfs.ext4 /dev/nvme7n1

mount /dev/nvme2n1 /sp/sp_dbb1

mount /dev/nvme5n1 /sp/sp_dbb2

mount /dev/nvme7n1 /sp/sp_archlog

# Edit /etc/fstab to ensure that mount points will be restored on reboot

vi /etc/fstab

#############################################

# Spectrum Protect Instance User Setup

#############################################

# Run the following as the "root" user.

# Add a Linux group and user to own the IBM Spectrum Protect instance.

# Set the password for this instance.

groupadd tsmsrvrs

useradd -g tsmsrvrs tsminst1

53
passwd tsminst1

# Ensure that this user can use the mount points previously created

chown -R tsminst1:tsmsrvrs /sp/

# From here, proceed with the IBM Spectrum Protect installation and setup

IBM Spectrum Protect, Extra-Small Amazon EC2 System, Ubuntu Linux: m5a.xlarge
Instance Type
#############################################

# Miscellaneous

#############################################

# Update apt packages, install tooling

apt update

apt upgrade -y

apt install scsitools -y

apt install dstat -y

apt install sysstat -y

apt install multipath-tools -y

apt install multipath-tools-boot -y

# Start the multipath daemon

touch /etc/multipath.conf

systemctl restart multipath-tools.service

# Set server hostname

hostnamectl set-hostname <desired hostname>

# Set vm.swappiness to 5 to match v4.1 xLinux blueprint recommendation

sysctl vm.swappiness=5

# Install the Korn shell

apt install ksh -y

# Reboot the system

#############################################

# Spectrum Protect Disk Setup (Ubuntu)

#############################################
# Run the following as the "root" user

# Install LVM for Ubuntu

apt install lvm2 -y

systemctl enable lvm2-lvmetad

systemctl start lvm2-lvmetad

# Make the required Spectrum Protect directories

mkdir /sp

mkdir /sp/tsminst1

mkdir /sp/sp_db1

mkdir /sp/sp_db2

mkdir /sp/sp_db3

mkdir /sp/sp_db4

mkdir /sp/sp_dbb1

mkdir /sp/sp_alog

mkdir /sp/sp_archlog

mkdir /sp/sp_cc

# Note that disk name values may be different from those seen here in actual
deployments

# Instance directory

# Change the value for "nvme*" to match your environment

mkfs.xfs /dev/nvme6n1

mount /dev/nvme6n1 /sp/tsminst1

# Cloud Cache

pvcreate /dev/nvme2n1

vgcreate sp_cc /dev/nvme2n1

lvcreate --stripesize 16 --extents 100%FREE --name sp_cc_vg sp_cc

mkfs.xfs /dev/mapper/sp_cc-sp_cc_vg

mount /dev/mapper/sp_cc-sp_cc_vg /sp/sp_cc

# DB and actlog.

# Change the value for "nvme*" to match your environment

pvcreate /dev/nvme3n1

pvcreate /dev/nvme6n1

pvcreate /dev/nvme7n1

55
pvcreate /dev/nvme8n1

pvcreate /dev/nvme9n1

vgcreate sp_db /dev/nvme3n1 /dev/nvme6n1 /dev/nvme7n1 /dev/nvme8n1 /dev/nvme9n1

lvcreate --stripes 5 --stripesize 16 --extents 16385 --name sp_db_db1 sp_db

lvcreate --stripes 5 --stripesize 16 --extents 16385 --name sp_db_db2 sp_db

lvcreate --stripes 5 --stripesize 16 --extents 16385 --name sp_db_db3 sp_db

lvcreate --stripes 5 --stripesize 16 --extents 16385 --name sp_db_db4 sp_db

lvcreate --stripes 5 --stripesize 16 --extents 16375 --name sp_db_alog sp_db

mkfs.ext4 /dev/mapper/sp_db-sp_db_db1

mkfs.ext4 /dev/mapper/sp_db-sp_db_db2

mkfs.ext4 /dev/mapper/sp_db-sp_db_db3

mkfs.ext4 /dev/mapper/sp_db-sp_db_db4

mkfs.ext4 /dev/mapper/sp_db-sp_db_alog

mount /dev/mapper/sp_db-sp_db_db1 /sp/sp_db1

mount /dev/mapper/sp_db-sp_db_db2 /sp/sp_db2

mount /dev/mapper/sp_db-sp_db_db3 /sp/sp_db3

mount /dev/mapper/sp_db-sp_db_db4 /sp/sp_db4

mount /dev/mapper/sp_db-sp_db_alog /sp/sp_alog

# DB backup and Archive log

pvcreate /dev/nvme3n1

vgcreate sp_dbb_vg /dev/nvme3n1

lvcreate --extents 50%FREE --name sp_dbb_dbb sp_dbb_vg

lvcreate --extents 100%FREE --name sp_dbb_archlog sp_dbb_vg

mkfs.ext4 /dev/mapper/sp_dbb_vg-sp_dbb_dbb

mkfs.ext4 /dev/mapper/sp_dbb_vg-sp_dbb_archlog

mount /dev/mapper/sp_dbb_vg-sp_dbb_dbb /sp/sp_dbb1

mount /dev/mapper/sp_dbb_vg-sp_dbb_archlog /sp/sp_archlog

# Edit /etc/fstab to ensure that mount points will be restored on reboot

vi /etc/fstab

#############################################

# Spectrum Protect Instance User Setup

#############################################

# Run the following as the "root" user.

# Add a Linux group and user to own the IBM Spectrum Protect instance.
# Set the password for this instance.

groupadd tsmsrvrs

useradd -g tsmsrvrs tsminst1

passwd tsminst1

# Ensure that this user can use the mount points previously created

chown -R tsminst1:tsmsrvrs /sp/

# From here, proceed with the IBM Spectrum Protect installation and setup

Disk Benchmarking
As a part of vetting each AWS test configuration outlined in this document, disk benchmark
tests were performed to validate the capability of the disk volumes underlying the IBM
Spectrum Protect database and cloud accelerator cache. From a database point of view,
this vetting was done to ensure that the volumes were sufficiently capable from an IOPS
perspective to support the 8 KiByte random mixed write and read workload that a busy
Blueprint-level system would demand. From a cloud cache standpoint, the vetting was
performed to ensure that overlapped 128-256 KiByte write and read throughput could
achieve a rate high enough such that the server’s bottleneck for IO would be at the
instance-to-object storage network level and not the disk level. The goal was to ensure that
the disk could perform at a rate such that the IBM Spectrum Protect server could utilize it
during overlapped ingest and be able to stress the network link layer simultaneously.
Disk benchmarking was performed by using the tsmdiskperf.pl Perl script, provided as a
part of the Blueprint configuration scripts package found on the IBM Spectrum Protect
Blueprints page (References [1]). Execution of the script was performed as follows:
perl tsmdiskperf.pl workload=stgpool fslist=directory_list
perl tsmdiskperf.pl workload=db fslist=directory_list

With a stgpool workload specification, the script drives a 256 KiByte IO pattern, whereas
with a db workload specification, the script drives 8 KiByte operations. For each directory
location provided as a value to the comma-separate fslist, a pair of IO processes is
created to perform writes and reads to test files that are generated in that directory.
Typical script output for a stgpool workload run resembles the following example:
======================================================================

: IBM Spectrum Protect disk performance test (Program version 3.1b)

: Workload type: stgpool

: Number of filesystems: 1

57
: Mode: readwrite

: Files to write per fs: 5

: File size: 2 GB

======================================================================

: Beginning I/O test.

: The test can take upwards of ten minutes, please be patient ...

: Starting write thread ID: 1 on filesystem /sp/sp_cc/1

: Starting read thread ID: 2 on filesystem /sp/sp_cc/1

: All threads are finished. Stopping iostat process with id 111519

===================================================================

: RESULTS:

: Devices reported on from output:

: dm-2

: Average R Throughput (KB/sec): 19473.92

: Average W Throughput (KB/sec): 19377.45

: Avg Combined Throughput (MB/sec): 37.94

: Max Combined Throughput (MB/sec): 160.57

: Average IOPS: 464.63

: Peak IOPS: 2154.57 at 11/10/2017 04:22:32

: Total elapsed time (seconds): 443

===================================================================

The value that was extracted for the purposes of comparison and validation for stgpool
workloads was Avg Combined Throughput (MB/sec). The goal was to determine the
largest aggregate average throughput for writes and reads to the accelerator cache disk
such that overlapped backup ingest and transfer to object storage will not be constrained
by disk capability.
When running the tool in db workload mode, output should appear similar to the following
example:
======================================================================

: IBM Spectrum Protect disk performance test (Program version 3.1b)

:
: Workload type: db

: Number of filesystems: 1

: Mode: readwrite

: Thread count per FS: 2

: I/Os per thread: 500000

: File size: 10 GB

======================================================================

: Creating files of 10 GB to simulate IBM Spectrum Protect DB.

: Issuing command ./ldeedee if=/dev/zero of=/sp/sp_db1/1/tsmdiskperf_1 bs=1024k

count=10240 dio=2 > /dev/null 2>&1

: Issuing command ./ldeedee if=/dev/zero of=/sp/sp_db1/1/tsmdiskperf_2 bs=1024k

count=10240 dio=2 > /dev/null 2>&1

: Beginning I/O test.

: The test can take upwards of ten minutes, please be patient ...

: All threads are finished. Stopping iostat process with id 111978

===================================================================

: RESULTS:

: Devices reported on from output:

: dm-6

: Average R Throughput (KB/sec): 12907.53

: Average W Throughput (KB/sec): 12707.15

: Avg Combined Throughput (MB/sec): 25.01

: Max Combined Throughput (MB/sec): 42.70

: Average IOPS: 3202.28

: Peak IOPS: 5465.86 at 11/10/2017 04:31:47

: Total elapsed time (seconds): 30

===================================================================

For the db workload tests, the Avg Combined Throughput (MB/sec) and Average
IOPS metrics are significant for evaluating database disk capability. Here, the small
random IOPS capability of the underlying disk that is used for the IBM Spectrum Protect
Db2 database is of interest.

59
To conduct measurements of your own, increase the number of write/read threads pairs
(and directories) by 1 for each test until the average throughput, the average IOPS, or both
stabilize (level off). Benchmark test results are provided here as a reference for those who
want to build systems resembling those laid out in this document and who want to validate
that their system is capable of supporting the described level of ingestion. For each graph,
the horizontal axis represents the quantity of write/read thread pairs (and the number of
directory locations used with fslist). For each successive bar to the right, the thread
count affecting the disk is increased by 2 (1 write thread, 1 read thread, and adding a
directory location). The vertical axis represents total average throughput in MiBytes/s.

AWS Large Configuration

250
Number of Write/Read Thread Pairs
200
Throughput (MiBytes/s)

150

100

50

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 5: AWS large configuration; database volume average throughput; 8 KiByte random
writes/reads
 35000
Number of Write/Read Thread Pairs
30000

25000

20000
IOPs

15000

10000

5000

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 6: AWS large configuration; database volume average IOPS; 8 KiByte random writes/reads

700
Number of Write/Read Thread Pairs
600
Throughput (MiBytes/s)

500

400

300

200

100

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 7: AWS large configuration; cloud cache volume average throughput; mixed 256 KiByte writes
and reads

61
AWS Medium Configuration

100
Number of Write/Read Thread Pairs
90

Throughput (MiBytes/s) 80
70
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 8: AWS medium configuration; database volume average throughput; 8 KiByte random
writes/reads

12000
Number of Write/Read Thread Pairs

10000

8000
IOPs

6000

4000

2000

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 9: AWS medium configuration; database volume average IOPS; 8 KiByte random writes/reads
 500
Number of Write/Read Thread Pairs
450
400

Throughput (MiBytes/s)
350
300
250
200
150
100
50
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 10: AWS medium configuration; cloud cache volume average throughput; mixed 256 KiByte
writes and reads

AWS Small Configuration

70
Number of Write/Read Thread Pairs
60
Throughput (MiBytes/s)

50

40

30

20

10

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Figure 11: AWS small configuration; database volume average throughput; 8 KiByte random
writes/reads

63
 8000
Number of Write/Read Thread Pairs
7000

6000

5000
IOPs

4000

3000

2000

1000

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Figure 12: AWS small configuration; database volume average IOPS; 8 KiByte random writes/reads

255
Number of Write/Read Thread Pairs
250
245
Throughput (MiBytes/s)

240
235
230
225
220
215
210
205
200
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Figure 13: AWS small configuration; cloud cache volume average throughput; mixed 256 KiByte
writes/reads

Object Storage Benchmarking

Another important step in validating the capability of an IBM Spectrum Protect in-the-cloud
solution is to benchmark the throughput of the bare metal server or compute instance to
the chosen target object storage system with a workload that is typical of IBM Spectrum
Protect. Ideally, any in-the-cloud IBM Spectrum Protect solution should be network bound
in terms of its connection to object storage. Post inline data deduplication, compression,
and encryption, the back-end ingestion rate over HTTPS should dictate an upper bound for
daily ingestion performance of a system.
To help facilitate this test activity in-house, a Java program was developed by the IBM
Spectrum Protect test team to emulate the behavior of the server’s use of the Amazon S3
Java API. The tool can be used to drive various backup and restore-type activities to object
storage, including direct HTTP PUT, GET, multipart file upload, and range-read restore
behavior with a variable number of threads. This tool, known as SPObjBench.jar, is
included with the Benchmarking package provided with this document.
Also included in the Benchmarking package is a Perl script, tsmobjperf.pl, which can
be used to automate execution of the SPObjBench.jar file with several thread counts to
measure ingest (PUT) and restore (GET) scalability.
On the normal ingestion path within the scope of a direct-to-cloud with accelerator cache
architecture, the IBM Spectrum Protect server attempts to upload up to 100 1 GB disk
container files from the accelerator cache in parallel by using multipart upload. Within a
production environment, this work would occur in conjunction with overlapped ingestion to
another set of container files within the same storage pool directory locations on
accelerator cache disk storage.
To attempt to isolate and gauge the instance-to-object storage ingestion capability of a
system, you can complete the following steps:
1. Populate a set of 10 1 GB files in a memory-mapped file system location to use as
source data for ingestion. The use of memory-mapped locations (such as tmpfs on
Linux) is preferred to eliminate source disk bottlenecks. For a Linux system with at
least 11 GB of free RAM, run the following commands:
mkdir /mnt/ramdisk
mount –t tmpfs –o size=11g tmpfs /mnt/ramdisk
for I in `seq 10`; do dd if=/dev/urandom
of=/mnt/ramdisk/file.$I bs=1048576 count=1024; done

2. To run a set of automated tests scaling from 1 to 100 threads, run the
tsmobjperf.pl tool by using the recently created RAM disk files as source files to
upload. If more threads are specified than files are present in the source list, the tool
completes a round-robin action over these source files. Because all activity is read-
only, using separate file handles from memory-mapped sources, multiple threads
sharing the same file is not a concern. To test with 1, 10, 20, 30, 40, 50, 60, 70, 80, 90,
and 100 threads, run the tool as follows, specifying the arguments as needed:
perl tsmobjperf.pl type=type endpoints=endpoint user="user"
pass="pass" bucket=bucket min=1 max=100 step=10 flist=
comma_delimited_source_files_list

where:

• type should be s3 for Amazon S3.

65
 • endpoint specifies a comma-delimited list of IP addresses or URLs for the
object storage endpoints. For Amazon S3, this should be a single private S3
endpoint that is accessed over HTTPS (for security), preferably with a VPC
endpoint set up.

• For S3, the user is a public key ID.

• For S3, the pass is the secret key for a user who has valid S3 credentials to
create buckets and PUT and GET objects in the region indicated by the
endpoint URL. These values align with those that are used to define an IBM
Spectrum Protect cloud-container storage pool, either via the Operations
Center or the command line.

• The bucket value should be an Amazon S3 bucket name that the credentialed
user has create/PUT/GET access to and that exists in the object storage
system. The min and max values should indicate the minimum and maximum
thread counts to test.

• The step value should indicate the increase in thread count from test to test.

• The flist parameter should include a comma-delimited list of source files to

be used for multipart upload. These files should be the same as those created
earlier in the memory-mapped file system.
The following example is for execution of an Amazon S3 based endpoint in the US
West (Oregon) Region, using 100 upload threads with an existing test bucket:
perl tsmobjperf.pl type=s3 endpoints=https://s3.us-west-
2.amazonaws.com user=”PUBLICKEYID” pass=”SECRETKEY”
bucket=testbucket min=1 max=100 step=10
flist=/mnt/ramdisk/file.1,/mnt/ramdisk/file.2,/mnt/ramdisk/file.
3,/mnt/ramdisk/file.4,/mnt/ramdisk/file.5,/mnt/ramdisk/file.6,/m
nt/ramdisk/file.7,/mnt/ramdisk/file.8,/mnt/ramdisk/file.9
,/mnt/ramdisk/file.10

Each thread count test (for 1, 10, 20, or more threads) uploads 10 x 1 GB objects per
thread. The previous example would result in a total of 5510 GB of data being stored to
the test bucket after all thread tests are completed. The tool does not remove objects
that are created. You must remove the objects manually after test completion.
Upon completion, the tool generates aggregate throughput metrics that can be used to
estimate practical instance-to-object storage performance rates for IBM Spectrum
Protect. Data is provided in comma-separated-value format (CSV) and the output of
the SPObjBench.jar tool can be inspected upon completion as well:
===================================================================

: IBM Spectrum Protect object storage test

: Test Mode: write

: Type: s3

: Endpoints: https:// s3.us-west-2.amazonaws.com

: User: PUBLICKEY

: Pass: SECRETKEY

: Test Bucket: testbucket

: Min Threads: 1

: Max Threads: 100

: Thread Step: 10

: File List:
/mnt/ramdisk/file.1,/mnt/ramdisk/file.2,/mnt/ramdisk/file.3,/mnt/ramdisk/file.4,
/mnt/ramdisk/file.5,/mnt/ramdisk/file.6,/mnt/ramdisk/file.7,/mnt/ramdisk/file.8,
/mnt/ramdisk/file.9 ,/mnt/ramdisk/file.10

: Using java: java

===================================================================

SPObjBench.jar output being captured to file: tsmobjperf.1540336631.out

===================================================================

: Test Results

Thread Count, Write Throughput (MB/s)

1, XXX

10, XXX

20, XXX

30, XXX

40, XXX

50, XXX

60, XXX

70, XXX

80, XXX

90, XXX

100, XXX

===================================================================

It can be beneficial to monitor network transmission rates externally from the tool, as
well, to validate the absolute throughput rate that is experienced to object storage over
the (Ethernet) network. The tool reports an aggregate rate that can include build-up
and tear-down overhead associated with the tool. Calculating an actual transmission
rate from the instance-to-object storage while the test is running can give an indication
of the throughput limits of the environment. On Linux, for example, the dstat utility
can be used to monitor several system metrics at once, including network interface
send and receive statistics, by using the basic command:

67
 % dstat

You did not select any stats, using -cdngy by default.

----total-cpu-usage---- -dsk/total- -net/total- ---paging-- ---system--

usr sys idl wai hiq siq| read writ| recv send| in out | int csw

0 0 100 0 0 0| 60B 2511B| 0 0 | 0 0 | 76 71

15 1 84 0 0 1| 0 24k|1674k 58M| 0 0 | 42k 2785

15 1 83 0 0 1| 0 0 |1838k 62M| 0 0 | 46k 2969

16 1 82 0 0 1| 0 0 |1832k 61M| 0 0 | 45k 3127

15 1 84 0 0 1| 0 0 |1753k 61M| 0 0 | 44k 2822

16 1 83 0 0 1| 0 0 |1811k 62M| 0 0 | 45k 3001

15 1 83 0 0 1| 0 0 |1778k 62M| 0 0 | 45k 3068

16 1 82 0 0 1| 0 0 |1870k 63M| 0 0 | 46k 3068

16 1 82 0 0 1| 0 0 |1933k 64M| 0 0 | 46k 3011

15 1 83 0 0 1| 0 0 |1834k 63M| 0 0 | 46k 2974

The dstat tool outputs a new line of metrics at a configured interval, much like the
standard iostat and netstat utilities. For the execution above, the net/total
send column is of greatest interest, here reported in MiBytes, as an indication of how
quickly data could be sent to the object storage endpoint from the server.

Instance and Object Storage: Navigating the

AWS Portal
When deploying a new IBM Spectrum Protect environment based on AWS, you must
navigate the web portal and create relevant cloud resources. Use the following steps as a
starting point for creating AWS resources. An important consideration when building an
IBM Spectrum Protect server in AWS is to correctly configure host servers and object
storage such that the private (internal AWS) network link between these resources is
efficient.
Begin by navigating to the AWS home page and sign in with the appropriate user
credentials.
AWS EC2 Instances
1. At the top left, click Services. In the Compute section, click EC2.

2. In the Create Instance section, click Launch Instance.

3. Select an appropriate and supported Amazon Machine Image (AMI), for example,
Ubuntu Server 16.04 LTS.

4. Select an instance type that is applicable to the desired IBM Spectrum Protect cloud
Blueprint size. Click Next: Configure Instance Details.

69
5. On the “Step 3: Configure Instance Details” page, customize the VPC, subnet, IAM
role, and other settings related to the instance. Click Next: Add Storage.
6. On the “Step 4: Add Storage” page, add the necessary IBM Spectrum Protect EBS
block disk appropriate to the desired cloud Blueprint size. Optionally, select the delete
on termination check box if EBS volumes should be deleted when the instance is
deleted.
Warning: Deletion of EBS volumes might not be desirable with IBM Spectrum Protect.
EBS volumes contain persistent IBM Spectrum Protect storage pool and database
data that might be required outside the lifetime of the instance. For example, it might
be necessary to terminate an EC2 instance and provision one of a different size or
model in some circumstances. Maintaining EBS disks with IBM Spectrum Protect data
separate from the instance will allow for greater instance flexibility.

7. Click Next: Add Tags. Add any instance-specific tags that you require. Then, click
Next: Configure Security Group.
8. On the “Step 6: Configure Security Group” page, assign the instance to a new or
existing Amazon EC2 security group. Security groups help to organize incoming or
outgoing protocol and port restriction rules for one or more instances in a group. Note
that outgoing TCP port 80 is required for HTTP communication to S3 object storage
and TCP port 443 is required for HTTPS. Click Review and Launch.
9. On the final “Step 7: Review Instance Launch” page, review the desired instance
settings. If the settings are correct, click Launch to direct Amazon to create the
instance and attached volumes.
AWS S3 Object Storage
1. At the top left, click Services. Then, in the Storage section, click S3.

2. Click Create bucket to create a bucket for use by IBM Spectrum Protect cloud-
container storage pools.

3. Specify a globally-unique name for the bucket and select an Amazon Region where
the bucket will be created. For optimal performance, the Amazon S3 bucket and
Amazon EC2 instance hosting IBM Spectrum Protect should be in the same Amazon

71
 Region. Click Next.

4. On the “Configure options” pane of the create bucket wizard, specify any further
options that you require.
Tips:
- Versioning is not recommended with IBM Spectrum Protect because this feature is
not used by cloud-container storage pools.
- If data privacy is a concern, object-level encryption is not necessary because
deduplicated extent-level encryption can be enabled on the IBM Spectrum Protect
cloud-container storage pool with little, if any, performance penalty.

Click Next.

5. On the “Set Permissions” page, adjust public bucket access, if desired. The preferred
method is to keep the default settings. Click Next.
 6. On the final “Review” page, review the desired bucket settings. If the settings are
correct, click Create bucket to direct Amazon to create the bucket within the desired
region.
When you access Amazon S3 object storage from an Amazon EC2 instance, be sure to use the
optimal S3 protocol endpoint. For a listing of available endpoints, see References [8]. If possible,
use the private endpoint that is geographically closest to the Amazon EC2 compute instance to
help optimize throughput.

73
 REFERENCES

[1] IBM Spectrum Protect Blueprints:

https://www.ibm.com/developerworks/community/wikis/home?lang=en#!/wiki/Tivoli%20Stor
age%20Manager/page/IBM%20Spectrum%20Protect%20Blueprints
[2] Amazon Web Services:
https://aws.amazon.com
[3] Amazon EC2, “Resizing an Amazon EBS-backed Instance”:
http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-resize.html
[4] Amazon EBS-Optimized Instances:
https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSOptimized.html
[5] Amazon EBS Volume Types:
https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSVolumeTypes.html
[6] AWS Direct Connect:
https://aws.amazon.com/directconnect/
[7] Overview - IBM Spectrum Protect Supported Operating Systems:
http://www.ibm.com/support/docview.wss?uid=swg21243309
[8] AWS Service Endpoints:
https://docs.aws.amazon.com/general/latest/gr/rande.html
[9] Amazon Nitro System:
https://aws.amazon.com/ec2/nitro/
[10] Announcing S3 Intelligent-Tiering:
https://aws.amazon.com/about-aws/whats-new/2018/11/s3-intelligent-tiering/
In-the-Cloud Deployment Guidelines with Amazon Web Services

Notices
This information was developed for products and services offered in the US. This material might be available from IBM in
other languages. However, you may be required to own a copy of the product or product version in that language in order to
access it.

IBM may not offer the products, services, or features discussed in this document in other countries. Consult your local IBM
representative for information on the products and services currently available in your area. Any reference to an IBM
product, program, or service is not intended to state or imply that only that IBM product, program, or service may be used.
Any functionally equivalent product, program, or service that does not infringe any IBM intellectual property right may be
used instead. However, it is the user's responsibility to evaluate and verify the operation of any non-IBM product, program,
or service.

IBM may have patents or pending patent applications covering subject matter described in this document. The furnishing of
this document does not grant you any license to these patents. You can send license inquiries, in writing, to:

IBM Director of Licensing

IBM Corporation
North Castle Drive, MD-NC119
Armonk, NY 10504-1785
US

For license inquiries regarding double-byte character set (DBCS) information, contact the IBM Intellectual Property
Department in your country or send inquiries, in writing, to:

Intellectual Property Licensing

Legal and Intellectual Property Law
IBM Japan Ltd.
19-21, Nihonbashi-Hakozakicho, Chuo-ku
Tokyo 103-8510, Japan

INTERNATIONAL BUSINESS MACHINES CORPORATION PROVIDES THIS PUBLICATION "AS IS" WITHOUT
WARRANTY OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Some
jurisdictions do not allow disclaimer of express or implied warranties in certain transactions, therefore, this statement may
not apply to you.

This information could include technical inaccuracies or typographical errors. Changes are periodically made to the
information herein; these changes will be incorporated in new editions of the publication. IBM may make improvements
and/or changes in the product(s) and/or the program(s) described in this publication at any time without notice.

Any references in this information to non-IBM websites are provided for convenience only and do not in any manner serve
as an endorsement of those websites. The materials at those websites are not part of the materials for this IBM product and
use of those websites is at your own risk.

IBM may use or distribute any of the information you provide in any way it believes appropriate without incurring any
obligation to you.

Licensees of this program who wish to have information about it for the purpose of enabling: (i) the exchange of information
between independently created programs and other programs (including this one) and (ii) the mutual use of the information
which has been exchanged, should contact:

IBM Director of Licensing

IBM Corporation
North Castle Drive, MD-NC119
Armonk, NY 10504-1785
US

Such information may be available, subject to appropriate terms and conditions, including in some cases, payment of a fee.

75
The licensed program described in this document and all licensed material available for it are provided by IBM under terms
of the IBM Customer Agreement, IBM International Program License Agreement or any equivalent agreement between us.

The performance data discussed herein is presented as derived under specific operating conditions. Actual results may
vary.

Information concerning non-IBM products was obtained from the suppliers of those products, their published
announcements or other publicly available sources. IBM has not tested those products and cannot confirm the accuracy of
performance, compatibility or any other claims related to non-IBM products. Questions on the capabilities of non-IBM
products should be addressed to the suppliers of those products.

This information is for planning purposes only. The information herein is subject to change before the products described
become available.

This information contains examples of data and reports used in daily business operations. To illustrate them as completely
as possible, the examples include the names of individuals, companies, brands, and products. All of these names are
fictitious and any similarity to actual people or business enterprises is entirely coincidental.

COPYRIGHT LICENSE:

This information contains sample application programs in source language, which illustrate programming techniques on
various operating platforms. You may copy, modify, and distribute these sample programs in any form without payment to
IBM, for the purposes of developing, using, marketing or distributing application programs conforming to the application
programming interface for the operating platform for which the sample programs are written. These examples have not been
thoroughly tested under all conditions. IBM, therefore, cannot guarantee or imply reliability, serviceability, or function of
these programs. The sample programs are provided "AS IS", without warranty of any kind. IBM shall not be liable for any
damages arising out of your use of the sample programs.

Trademarks
IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International Business Machines Corp.,
registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other
companies. A current list of IBM trademarks is available on the web at "Copyright and trademark information" at
www.ibm.com/legal/copytrade.shtml.

Intel and Itanium are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and
other countries.

Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both.

Java™ and all Java-based trademarks and logos are trademarks or registered trademarks of Oracle and/or its affiliates.

VMware is a registered trademark of VMware, Inc. or its subsidiaries in the United States and/or other jurisdictions.