Programming Intel® QuickAssist Technology Hardware Accelerators for Optimal Performance

Programming Intel® QuickAssist
Technology Hardware Accelerators for
Optimal Performance
White Paper
April 2015
Document Number: 332125-002

You may not use or facilitate the use of this document in connection with any infringement or other legal analysis concerning Intel products
described herein. You agree to grant Intel a non-exclusive, royalty-free license to any patent claim thereafter drafted which includes subject
matter disclosed herein.
No license (express or implied, by estoppel or otherwise) to any intellectual property rights is granted by this document.
All information provided here is subject to change without notice. Contact your Intel representative to obtain the latest Intel product
specifications and roadmaps.
The products described may contain design defects or errors known as errata which may cause the product to deviate from published
specifications. Current characterized errata are available on request.
Copies of documents which have an order number and are referenced in this document may be obtained by calling 1-800-548-4725 or by
visiting: http://www.intel.com/design/literature.htm.
Intel technologies’ features and benefits depend on system configuration and may require enabled hardware, software or service activation.
Learn more at http://www.intel.com/ or from the OEM or retailer.
No computer system can be absolutely secure.
Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will
affect actual performance. Consult other sources of information to evaluate performance as you consider your purchase. For more complete
information about performance and benchmark results, visit http://www.intel.com/performance.
Intel, Intel Core, Xeon, and the Intel logo are trademarks of Intel Corporation in the U.S. and/or other countries.
*Other names and brands may be claimed as the property of others.
© 2015, Intel Corporation.
Programming Intel®
QuickAssist Technology Hardware Accelerators
White Paper April 2015
2 Document Number: 332125-002

Contents
1.0 Introduction................................................................................................................................................. 6
1.1 Intended Audience..........................................................................................................................................6
1.2 Overview of this Document.........................................................................................................................6
1.3 Terminology.......................................................................................................................................................7
1.4 Reference Documents...................................................................................................................................8
2.0 Intel® QuickAssist Technology Overview............................................................................................10
2.1 Parallel Compression Benchmark..........................................................................................................11
3.0 Background................................................................................................................................................13
4.0 Accelerator Overview ..............................................................................................................................14
4.1 Asynchronous Hardware Interface........................................................................................................14
4.2 Multiple Parallel Engines............................................................................................................................16
4.3 Direct Memory Access.................................................................................................................................17
5.0 Parallelizability.........................................................................................................................................19
5.1 Symmetric Cryptography...........................................................................................................................19
5.2 Public Key Cryptography ...........................................................................................................................20
5.3 Compression ...................................................................................................................................................20
6.0 Programming Model................................................................................................................................22
6.1 Software, Multi-Core and Multi-Threading........................................................................................22
6.2 Hardware, Synchronous, Single Threaded.........................................................................................25
6.3 Hardware, Synchronous, Multi-Threading .........................................................................................27
6.4 Hardware, Asynchronous...........................................................................................................................29
6.5 Multi-Core and Hyper Threading............................................................................................................30
6.6 Multi-Socket and NUMA.............................................................................................................................30
6.7 Batch Submissions........................................................................................................................................30
6.8 Polling vs. Interrupts....................................................................................................................................30
6.9 Programming Model Summary...............................................................................................................32
7.0 Memory Model..........................................................................................................................................33
8.0 Methodology .............................................................................................................................................34
8.1 Benchmark Application ..............................................................................................................................34
8.2 Compression Framework...........................................................................................................................34
8.2.1 Chunks.............................................................................................................................................36
8.2.2 Buffer Sizes....................................................................................................................................37
8.3 Providers...........................................................................................................................................................37
8.3.1 zlib Provider..................................................................................................................................38
Programming Intel®
April 2015 White Paper
Document Number: 332125-002 3

8.3.2 qat Provider...................................................................................................................................38
8.4 Inputs..................................................................................................................................................................40
8.5 Outputs..............................................................................................................................................................42
8.5.1 Compression Ratio.....................................................................................................................42
8.5.2 Compression Elapsed Time ...................................................................................................42
8.5.3 Throughput ...................................................................................................................................43
8.5.4 Compression CPU Time...........................................................................................................43
8.6 Platform Description....................................................................................................................................44
9.0 Related Work.............................................................................................................................................45
10.0 Future Work...............................................................................................................................................46
11.0 Summary of Recommendations............................................................................................................47
Figures
Figure 1. Parallel Compression Software Stack ....................................................................................................12
Figure 2. Multi-Socket and NUMA...............................................................................................................................18
Figure 3. Compression Ratio vs. Chunk Size...........................................................................................................21
Figure 4. Multi-Threading and Multi-Core...............................................................................................................23
Figure 5. Impact of Multi-Threading on zlib Provider (SW Only, No Acceleration, No HT)................24
Figure 6. Calling Hardware Synchronously from a Single Thread ................................................................25
Figure 7. Time Taken to Compress 1 GB using Software and Hardware ...................................................26
Figure 8. Calling Hardware from Multiple Threads or Asynchronously......................................................27
Figure 9. Hardware Offload from Multiple Threads Synchronously............................................................28
Figure 10. Calling Hardware Asynchronously from a Single Thread..............................................................29
Figure 11. Comparison of Different Programming Models.................................................................................32
Figure 12. Multi-Threading and Segments ................................................................................................................35
Figure 13. Compressing in Chunks................................................................................................................................37
Figure 14. Flowchart for QAT Provider - Compression Function.....................................................................38
Figure 15. Flowchart for QAT Provider - Compression Callback Function..................................................39
Tables
Table 1. Terminology.........................................................................................................................................................7
Table 2. Reference Documents..................................................................................................................................... 8
Table 3. Popular Cryptographic and Compression Frameworks and Libraries.....................................10
Table 4. Number of Engines in Intel® Communications Chipset 8955......................................................16
Table 5. Affinity Mask......................................................................................................................................................41
Programming Intel®

Revision History
Date Revision Description
April 2015 002 Updated Table 1 Terminology.
March 2015 001 Initial release.
§
Programming Intel®

Introduction
1.0 Introduction
Intel®
QuickAssist Technology allows compute-intensive workloads, specifically
cryptography and compression, to be offloaded from the CPU core onto dedicated
hardware accelerators. Intel is working to enable software applications and frameworks
to take advantage of this offload seamlessly and transparently, achieving optimal
performance with minimal changes to the application. Integrating the functionality is
relatively straightforward. Doing so in a way which is optimal from a performance
perspective can be more challenging, however, due to differences in the programming
and memory models between the hardware and traditional software applications. This
document discusses some techniques for overcoming these differences. Using a
worked example of a sample application which performs compression, it specifically
demonstrates how to apply some of these techniques to perform parallelization under
the hood of a synchronous, user space API. The results of this benchmark
demonstrates that an approximate 95% reduction was achieved in elapsed time and
CPU time required to compress a 1 GB file relative to a software implementation based
on zlib, while achieving approximately the same compression ratio.
The document is intended to arm developers with the knowledge they need to be able
to identify opportunities for parallelism in various applications; to identify the best layer
at which to exploit this parallelism; and to program hardware based on Intel®
QuickAssist Technology to achieve optimal performance.
1.1 Intended Audience
The intended audience includes software designers and developers programming
applications using Intel®
QuickAssist Technology who are using the Intel®
QuickAssist
API, or who are programming adapters which allow Intel®
QuickAssist Technology to be
“plugged in” to other cryptographic and/or compression libraries/frameworks.
1.2 Overview of this Document
The document is structured as follows:
• Section 2.0 Intel® QuickAssist Technology Overview presents an overview of the
Intel®
QuickAssist Technology, and the example of an application which performs
compression is introduced.
• In Section 3.0 Background, some background on cryptography and compression is
provided including why these workloads are important.
• In Section 4.0 Accelerator Overview, an overview of the accelerator hardware is
given, focusing on the inherently asynchronous nature of the hardware interface,
Programming Intel®

Introduction
and how the hardware consists of multiple parallel engines, all of which need to be
kept busy in order to achieve the maximum benefit. Also discussed is the memory
model used by the hardware.
• Section 5.0 Parallelizability discusses parallelism in general, and specifically the
parallelism inherent in cryptography and compression.
• Section 6.0 Programming Model discusses how software can deal with both the
asynchronous nature of the hardware with the need to keep multiple parallel
engines busy.
• Section 7.0 Memory Model discusses how software can deal with the hardware’s
memory requirements is discussed.
• Section 8.0 Methodology describes a simple framework and application which has
been developed to demonstrate some of the principles described in the earlier
sections, and to show the results of that benchmark.
• Section 9.0 Related Work gives a reference to some additional key design decisions
that should be considered and other techniques that may be employed.
• Section 10.0 Future Work states some examples of possible future work.
• Finally, the recommendations from this document are summarized in Section 11.0.
1.3 Terminology
Table 1. Terminology
Term Description
AES Advanced Encryption Standard
API Application Programming Interface
CNG Cryptography API: Next Generation. This is a cryptographic API and framework
that is part of Microsoft* Windows*.
CPU Central Processing Unit. A CPU can contain multiple cores. A platform can
contain multiple CPUs, each in its own socket.
CTR Counter, a mode of block ciphers
DMA Direct Memory Access
DRAM Dynamic Random Access Memory
ECB Electronic Code Book, a mode of block ciphers
GB Gigabyte
Gbps Gigabits per second
HMAC Keyed-Hash Message Authentication Code, as defined by IETF RFC2104
Programming Intel®

Introduction
Term Description
IPsec IP Security, a cryptographic protocol
IV Initial Value (or Initialization Vector), an input to block ciphers in certain modes
KB Kilobyte
MB Megabyte
OpenSSL* An open source project that implements the Secure Sockets Layer (SSL) and
Transport Layer Security (TLS) protocols, as well as a full-strength general
purpose cryptography library.
SHA1 Secure Hash Algorithm, first generation, with output size 160 bits
SHA256 Secure Hash Algorithm, second generation, with output size 256 bits
SSL Secure Sockets Layer, a cryptographic protocol
TLS Transport Layer Security, a cryptographic protocol
XEX Xor-Encrypt-Xor, a mode of block ciphers
XTS XEX-based tweaked-codebook mode with ciphertext stealing, a mode of block
ciphers
zlib A software library used for data compression.
1.4 Reference Documents
Table 2. Reference Documents
Document Document
No./Location
[1] Cisco, "Cisco Global Cloud Index: Forecast and
Methodology, 2013–2018" 2014.
http://www.cisco.com/c/en/us/solutions/coll
ateral/service-provider/global-cloud-index-
gci/Cloud_Index_White_Paper.pdf
[2] ZDNet, "Microsoft to encrypt network traffic amid NSA
datacenter link tapping claims" 2013.
http://www.zdnet.com/microsoft-to-encrypt-
network-traffic-amid-nsa-datacenter-link-
tapping-claims-7000023687/
[3] Vinodh Gopal, Jim Guilford, Wajdi Feghali, Erdinc
Ozturk, Gil Wolrich, Kirk Yap, Sean Gulley, Martin Dixon,
"Cryptographic Performance on the 2nd Generation Intel®
Core™ processor family" January 2011.
http://www.intel.com/content/dam/www/pub
lic/us/en/documents/white-papers/2nd-gen-
core-cryptographic-performance-paper.pdf
[4] A. Polyakov, "openssl-1.0.2-beta3." http://permalink.gmane.org/gmane.comp.enc
ryption.openssl.announce/127
Programming Intel®

Introduction
Document Document
No./Location
[5] Schmidt, Douglas C. & Cranor, Charles D., "Half-
Sync/Half-Async: An Architectural Pattern for Efficient and
Well-structured Concurrent I/O," in Pattern Languages of
Program Design, 1995.
http://www.cs.wustl.edu/~schmidt/PDF/PLoP
-95.pdf
[6] "Parallel Compression Benchmark." -
[7] "RtlCompressBuffer function." http://msdn.microsoft.com/en-
us/library/windows/hardware/ff552127%28
v=vs.85%29.aspx
[8] "The Calgary Corpus." http://corpus.canterbury.ac.nz/resources/cal
gary.tar.gz
[9] " Intel®
QuickAssist Technology Performance
Optimization Guide."
https://01.org/sites/default/files/page/33068
7_qat_perf_opt_guide_rev_1.0.pdf
§
Programming Intel®

Intel® QuickAssist Technology Overview
2.0 Intel® QuickAssist Technology Overview
Intel®
QuickAssist Technology offers acceleration for two key compute intensive
workloads, namely cryptography and compression. To help customer applications take
advantage of this new acceleration technology, Intel is enabling certain key software
frameworks and applications.
Fortunately, most software applications that use cryptography and compression today
do so via a handful of well-established frameworks and libraries. Some of the most
commonly used such frameworks are listed in Table 1 Popular Cryptographic and
Compression Frameworks and Libraries. Many of these frameworks are extensible, i.e.,
they allow other implementations─including hardware-based implementations─to be
“plugged in” under the hood. This enables applications to take advantage of the
acceleration seamlessly and transparently.
Table 3. Popular Cryptographic and Compression Frameworks and Libraries
Framework Operating
Environment
Services Offered Plug In Support? Asynchronous
Support
OpenSSL*
libcrypto
Linux* user space,
also portable to
other operating
systems
Symmetric crypto,
public key crypto
Yes Work In
Progress1
CNG Windows* Symmetric crypto,
public key crypto
Yes No
Linux Kernel
Cryptographic
Framework (aka
scatterlist API)
Linux kernel Symmetric crypto Yes Yes (for
symmetric
crypto)
zlib Most operating
systems
supported
Compression No No
Integrating a hardware accelerator underneath these frameworks so that it is functional
is generally relatively straightforward. It involves adapting between the existing APIs
used by the frameworks/applications and those provided by Intel®
QuickAssist
Technology.
Integrating a hardware accelerator such that it achieves optimal performance can be
more challenging. Many software applications are single-threaded and use
synchronous APIs to access cryptography and/or compression. They also operate in
user space using virtually addressed, virtually contiguous, paged memory. As shown
1
At the time of writing, asynchronous support in OpenSSL is available on a
development branch on http://www.openssl.org.
Programming Intel®

below, the hardware interface is inherently asynchronous, has parallel engines which
need to be kept busy, and because it performs DMA, it requires a physically addressed,
physically contiguous, pinned memory. Adapting between these programming and
memory models can consume a lot of CPU cycles, reducing or even negating2
the
benefits of offloading the workload in the first place.
Integrating Intel®
QuickAssist Technology into these frameworks for optimal
performance requires that developers be familiar with how the hardware operates, and
how to program to it to achieve optimal performance. This document attempts to
describe some of the key concepts, and through a worked example, demonstrates
some of the techniques that can be used to achieve optimal performance. It also
references other documents where key performance-related learnings are captured.
2.1 Parallel Compression Benchmark
Throughout the document, the example of an application which performs compression
is used. As described in Section 8.0 Methodology, a simple compression framework was
developed. Similar to some of the common cryptographic and compression
frameworks described in Table 3 above, the framework allows different
implementations, called “providers”, to be plugged in. Two providers were
implemented, to allow for comparison: one a software implementation based on zlib
and the other a hardware implementation using Intel®
QuickAssist Technology. The
hardware provider (called “qat”) can be configured via parameters to use the hardware
either synchronously or asynchronously, allowing for parallelism. The framework also
supports multi-threading, which is another means to achieve parallelism. Also
developed was a simple command-line application called parcomp, which uses this
framework to perform parallel compression.
This benchmark acts as an example of how to overcome the performance challenges
above:
• The application runs in user space.
• It operates on data buffers in virtual, paged memory.
• It offers a synchronous API to the application.
This document describes how, within these constraints, a throughput was achieved of
up to over 20 Gbps of compression, cutting the time taken to compress a 1 GB file to
approximately 0.7 seconds (compared to approximately 19 seconds using zlib) while
achieving approximately the same compression ratio, and cutting the total CPU cycles
consumed by a similar ratio.
2
For example, for cipher algorithms such as AES which are relatively cheap in software
using Intel®
Advanced Encryption Standard New Instructions (Intel®
AES-NI), and for
small buffer sizes (e.g., less than a few hundred bytes), offloading may actually
consume more cycles than performing the operation in software.
Programming Intel®

The software stack is illustrated in Figure 1. Parallel Compression Software Stack. For
details of the software stack, see Section 8.0 Methodology.
Figure 1. Parallel Compression Software Stack
§
Programming Intel®

Background
3.0 Background
The volume of data flowing into and out of the world’s data centers continues to grow
[1]. More and more of this data is being encrypted before it is sent across the network
or stored on disk [2] to provide confidentiality. This in turn requires secret session keys
to be exchanged between the parties by using public key cryptography algorithms and
key agreement protocols. In addition, much of this data is compressed before being
encrypted in order to reduce storage and network bandwidth requirements. Public key
cryptography, data compression, and encryption are all compute-intensive workloads;
performing these algorithms in software can consume significant CPU cycles on the
data center servers and reduce the number of cycles available to run the applications
and services for which customers pay.
Intel®
QuickAssist Technology allows these compute-intensive workloads to be
offloaded from the CPU to hardware accelerators, which are more efficient in terms of
cost and power than general purpose CPUs for these specific workloads. A server
platform that performs a lot of cryptography and/or compression, and which includes
Intel®
QuickAssist Technology, can therefore potentially achieve higher application
performance in the same cost/power envelope─or can achieve the same performance
in a lower cost/power envelope─than a platform that uses software only to perform
these workloads.
§
Programming Intel®

Accelerator Overview
4.0 Accelerator Overview
This section describes at a high level the hardware implementing Intel®
QuickAssist
Technology. The main points from a software application developer’s perspective are
as follows:
• The hardware interface is inherently asynchronous.
• The hardware consists of multiple parallel engines, all of which need to be kept
busy in order to achieve the maximum throughput (and therefore benefit) of the
hardware.
• Data sent to the accelerator must be stored in physically contiguous, pinned
memory (or multiple such regions, described by a scatter-gather list).
Each of these aspects are described below in more detail. In later sections, how best to
develop software to take account of these aspects of the accelerator is described.
4.1 Asynchronous Hardware Interface
At the lowest level, the hardware accelerators that implement the Intel®
QuickAssist
Technology present a message-based request/response interface to the CPU. This is an
inherently asynchronous interface, i.e., the requested data transformation is not
complete immediately on returning from sending the request, but rather some time
later when the response is received. Contrast this with the sort of synchronous interface
that traditional software implementations provide, where the transformed data is
available as soon as the called function returns.
The operation proceeds as follows:
• The host software creates a request descriptor containing all the information
required by the hardware to perform the job, and writes it onto a request ring (or
queue) in the DRAM. It then writes to a doorbell register to inform hardware that
the request is ready to be processed. At this point, the host software is free to do
some other work.
• The accelerator fetches the request from the ring, and then fetches (from the
DRAM, via the DMA) all of the data required to process the request. It performs the
requested transformation (e.g., encryption), and then writes the resulting
transformed data back to the DRAM. Finally, it writes the response descriptor to the
response ring, and writes to a doorbell to indicate to the host software that the
response is ready to be processed.
• This doorbell can be configured to generate an interrupt to the host software, or
the host software can poll the response ring to see when the response is ready. The
host software then processes the response. Depending on the application, this
Programming Intel®

might involve sending a now-decrypted packet or SSL record to the protocol stack
to be processed, or writing the now-compressed data to an output buffer or file,
etc.
The above message-based request/response interface is abstracted by an application
programming interface (API) which is called the Intel®
QuickAssist Technology API. All of
the data required in the request descriptor is passed as parameters on this API.
To handle the asynchronous nature of the hardware, the concept of a callback function
is used. The programmer registers this callback function, which processes the response.
Depending on whether the application wishes to use an interrupt-driven or polled
model, the callback function is invoked as follows:
• If interrupts are enabled on the response rings, then the callback function will be
invoked in the context of an interrupt handler. In the Linux kernel, for example, this
is invoked in the context of a “bottom half”.
• Otherwise, the user must call a defined polling function to have responses
processed (see Section 6.8 Polling vs. Interrupts for more information on when this
should be called). This checks for responses on the specified response rings, and
for each response, calls the callback function. It then returns control to the calling
function.
Note: The programming for an asynchronous API like this typically involves separating the
processing of a job into two phases: what happens before the operation is invoked (the
pre-processing) and what happens afterwards (the post-processing). The callback
function is written to implement the post-processing, or to cause it to be invoked.
Synchronous semantics can also be implemented on top of an asynchronous interface,
as follows:
• If interrupts are being used, then the calling thread can pend on some
synchronization primitive (e.g., a semaphore), and the callback function can post to
this to unblock the thread.
• If polling is being used, then the function can periodically poll until the response is
available and has been processed; or a separate thread can do the polling and use a
synchronization primitive as described for the interrupt case. As above, see Section
6.8 Polling vs. Interrupts for more information on when this should be called.
Note: The Intel®
QuickAssist Technology API supports both synchronous and asynchronous
calling semantics. If no callback function is specified, then synchronous semantics are
assumed.
Programming Intel®

4.2 Multiple Parallel Engines
In order to achieve a desired level of performance, implementations of the Intel®
QuickAssist Technology include multiple hardware accelerators, or engines, which can
operate in parallel. Table 4. Number of Engines in Intel® Communications Chipset 8955
shows the number of engines for each workload that exist in Intel’s current generation
product.
Table 4. Number of Engines in Intel® Communications Chipset 8955
Workload Number of Parallel Engines per Device
Compression 6
Symmetric Cryptography 12
Public Key Cryptography 30
In order to achieve the full advertised throughput of the device, all of the underlying
engines must be kept busy. For example, to achieve the maximum compression
throughput, the software needs to keep six engines busy; to achieve the full
approximately 41K operations/second of RSA2K decryption, the software needs to
keep all 30 public key cryptography engines busy.
Note: The software does not need to load balance requests across these engines; this is
handled in hardware.
In order to keep all of these engines busy, there should be at least as many jobs
outstanding as there are hardware engines in the device. In fact, as the data later in this
document will show, it is generally required to have more jobs outstanding than the
number of engines, because even within these engines there is some pipelining, with
different jobs in different stages of the pipeline (e.g., fetching uncompressed data from
the DRAM, compressing data, writing compressed data to the DRAM). The data below in
fact shows exactly this: for compression, the throughput of the device plateaus when
there are 18 jobs outstanding (six engines, each operating on three jobs simultaneously
– see Figure 10. Calling Hardware Asynchronously from a Single Thread, later in this
document).
To be able to perform multiple jobs in parallel, the jobs need to be independent. There
can be no data dependency between the output of one job and the input of another.
See Section 5.0 Parallelizability for a discussion on parallelism.
For those cases where parallelism is possible, there are various mechanisms software
applications can use to keep the hardware busy. The preferred mechanisms will be
described in Section 6.0 Programming Model.
Programming Intel®

4.3 Direct Memory Access
Intel®
QuickAssist Technology is implemented in the hardware as a device which uses
Direct Memory Access (DMA) to access data in the DRAM. As such, the data to be
operated on must be in the DMA-able memory. In the current generation of hardware,
this means that the data must be stored in pages which are pinned, and the pages be
physically contiguous. Alternatively, software can pass data in multiple such regions,
described by a scatter-gather list.
An additional consideration in the multi-processor (MP) or multi-socket platforms is the
locality of both the accelerator device, and of the data in memory, to the processor on
which the software is running. Many servers consist of multiple processors or CPU
sockets, each with its own memory controller and local DRAM. Such systems are called
Non-Uniform Memory Architecture (NUMA) systems. When a device performs DMA to
read from or write to memory, the time taken to complete that operation will depend
on whether the memory is local to the socket to which the device is physically
connected, or is located in a remote socket. This is because requests to and responses
from remote memory incur additional latency across the QPI bus which connects the
sockets.
Figure 2. Multi-Socket and NUMA shows the case of a platform in which there are two
processors (sockets), each with its own local memory controller, and a single Intel®
QuickAssist Technology-enabled PCIe* card connected via the PCIe to Processor #0. If
software running on Processor #1, or with data located in DRAM #1, wishes to perform
acceleration, then the request, response, and data all have to incur additional latency
across the QPI bus. The impact of this additional latency to the performance of an
application is described later.
Programming Intel®

Figure 2. Multi-Socket and NUMA
§
Programming Intel®

Parallelizability
5.0 Parallelizability
As noted earlier, in order to take advantage of the parallel engines in the hardware, they
need to be operating on independent data. Some workloads and applications are easily
parallelizable; these are sometimes called “embarrassingly parallel”. In other cases, it is
not as easy, typically because of data dependencies.
Note: As described later, software implementations can also take advantage of
parallelizability. For example, they can distribute the work across multiple
cores/threads, or operate on multiple data with a single instruction (SIMD).
Another consideration is the granularity (size) of the parallelizable task. Creating a
request and sending it to hardware, and then processing the response later, consumes
cycles, called the “cost of offload”. To make offload worthwhile, the cost of offload
should be less than the cost of simply performing the task in software.
Consider each of the three workloads in turn in terms of their parallelizability, and the
granularity thereof.
5.1 Symmetric Cryptography
For block ciphers (e.g., AES), certain modes (e.g., ECB, CTR) are parallelizable in both the
encrypt and decrypt directions, meaning that each individual block (16 bytes, in the
case of AES) can be encrypted or decrypted independently of the others. Other modes
are parallelizable only in one direction, e.g., CBC is parallelizable in the decrypt
direction only, because of a data dependency in the encrypt direction, whereby the
output of the encryption of a block is used as input to the next block.
In terms of granularity, even for those modes/directions which are parallelizable,
offloading individual 16B blocks is almost certain to cost more in terms of “cost of
offload” than the job being offloaded. This is especially true for algorithms such as AES,
where the Intel®
AES-NI instruction allows this to be implemented on the CPU in a small
number of cycles. However, given a sufficiently large buffer to be encrypted or
decrypted, and a mode/direction which is parallelizable, it would be possible to break
this up into smaller chunks (but which are sufficiently large to be worthwhile
offloading), offload them as separate independent jobs, and then reassemble the
outputs into a single buffer.
Moving up the software stack, some cryptographic protocols allow parallelism, while
others do not. For example, in versions of TLS/SSL prior to 1.1, when using AES in the
CBC mode, multiple records from a single SSL session cannot be encrypted in parallel,
because the CBC residue of one record is used as the IV for the next record, which again
requires the implementation to be serialized. In this case, to achieve parallelism, one
would have to operate on records from different SSL connections/sessions; this is
Programming Intel®

Parallelizability
feasible assuming the application’s programming model allows this (see Section 6.0
Programming Model). In contrast, for TLS 1.1 and beyond, as well as in IPsec, the IV is
explicit in each record (or packet, for IPsec), which breaks the dependency between
records and allows multiple records from the same session to be encrypted in parallel.
Note: As noted earlier, software implementations can also take advantage of this kind of
parallelizability. For example, as described in reference [3], the AESENC instruction
(part of the Intel®
AES-NI instruction set), which does one round of AES encryption, has
a latency of several cycles, but a throughput of one block per cycle. In the case of the
CBC-encrypt, one cannot start encrypting a block until the previous block has been
encrypted, which leaves the AES engine in the CPU core under-utilized; performance is
limited by the latency, rather than the throughput. However, if encrypting multiple
independent buffers in parallel, the data dependencies can be broken and ideal
performance is limited only by the throughput. Intel has worked with the OpenSSL
Software Foundation to take advantage of this in OpenSSL* [4]. When an application
writes a data buffer to an SSL socket (via the function ssl3_write_bytes), if certain
constraints are met (e.g., the buffer is sufficiently large, the cipher suite is AES-128-
CBC+HMAC-SHA1 or -SHA256 and uses an explicit IV, etc.), then the implementation
will create multiple independent SSL/TLS records and perform the encryption for up to
eight such records in parallel.
For storage applications, these typically operate on disk blocks of 4 KB. They also tend
to use cipher modes such as XTS, in which each disk block can be encrypted and
decrypted independently, which allows for parallel processing of blocks.
5.2 Public Key Cryptography
Public key cryptography is not inherently parallelizable; each individual request to
perform an RSA decryption or other primitive operation cannot be parallelized. As in
the case of symmetric crypto, one must rely on the application to submit multiple such
independent jobs in parallel.
5.3 Compression
Compression does lend itself to being parallelized, albeit this involves a tradeoff of
some compression ratio. A sufficiently large input buffer can be broken into multiple
segments or chunks, each of which can be compressed “statelessly” (independently),
each yielding a complete deflate block that can also be decompressed independently.
Some compression opportunities may be lost in doing this (i.e., the compression cannot
take advantage of patterns repeated in different segments), but for sufficiently large
chunks, this is not typically significant. The parcomp application described in this
document, and shown in Figure 3. Compression Ratio vs. Chunk Size, shows that using
Intel®
QuickAssist Technology, a chunk size of 64 KB gives an optimal compression
ratio. For software (using zlib), larger buffer sizes give marginally better compression
ratio, but above 64 KB, the gains are modest.
Note: The compression ratio is heavily dependent on the data being compressed. This data
was gathered using the Calgary corpus, as described in Section 8.0 Methodology.
Programming Intel®

Parallelizability
Figure 3. Compression Ratio vs. Chunk Size
The main points from this are (at least for the Calgary corpus – see Section 8.0
Methodology):
• If the buffer to be compressed is sufficiently large (e.g., in the order of MB or larger),
then it is easily parallelized by breaking it into chunks of about 64 KB, with minimal
loss in compression ratio.
• Using Intel®
QuickAssist Technology, the optimal buffer size to offload is found to
be around 64-128 KB. Larger buffers (e.g., >128 KB) can yield sub-optimal Huffman
trees, leading to a poorer compression ratio.
§
Programming Intel®

Programming Model
6.0 Programming Model
As noted above, in order to achieve the maximum throughput from the hardware, it is
necessary to keep its multiple parallel engines busy, which in turn requires to have
multiple jobs outstanding at one time. This is analogous to the way in which a multi-
core CPU requires multiple software threads (or processes) to keep the multiple cores
busy.
To illustrate this, the next section briefly describes how to optimize a software
application for multi-core. Subsequent sections describe how to optimize for hardware
offload using the Intel®
QuickAssist Technology.
6.1 Software, Multi-Core and Multi-Threading
A single-threaded application does not run any faster on a multi-core CPU than it does
on a single-core CPU; it just ends up under-utilizing the CPU. In order to make it run
faster, the application needs to be modified to make it multi-threaded.
Note: Clearly, other software can be run on the other cores. In fact, many computers rely on
this: individual applications are single-threaded, but the operating system schedules
different processes to run on different cores, so that the system feels responsive
overall. However, to make individual applications run faster, multi-threading is needed.
Consider the example of the sample application which compresses files. The basic
steps required in such a program are:
• Read the uncompressed file from disk to memory.
• Compress the data in memory.
• Write the compressed data from memory to disk.
For larger files, the bulk of the CPU cycles will be spent performing the compression.
(Note that applications like this, where the bulk of the cycles are spent performing the
workload which can be offloaded using Intel®
QuickAssist Technology, will clearly
benefit most from the technology.) Figure 4 (a) shows what this looks like in terms of
what the CPU utilization on a multi-core CPU. The green bars represents cycles spent
reading and writing the file to/from disk, processing command line parameters, etc.; the
blue bar represents the time spent actually compressing the data (e.g., in zlib). This
single-threaded application does not run any faster on a multi-core CPU; rather, the
remaining cores simply remain unused.
Programming Intel®

Programming Model
Figure 4. Multi-Threading and Multi-Core
To speed up the compression time on a multi-core CPU, the application can be
reworked to be multi-threaded. As described earlier, the file can be broken into N
segments, and these can be compressed (statelessly) in parallel in separate software
threads, which the operating system will schedule on the different cores, as shown in
Figure 4 (b). The main thread then waits for all the other threads to complete, and
reassembles the output into a single compressed file.
By spreading the load across N cores, the compression phase now operates up to N
times faster. The total CPU cycles increase marginally, as the software now has some
extra work to do in setting up, tearing down, and context switching between the
threads, reassembling the output, etc. This is represented by the red bars in Figure 4.
Again, for the case where there is a lot of data to be compressed, this cost is relatively
small. For smaller files, however, the benefit of parallelizing the compression will be
more modest.
To see what this looks like in practice, consider the parallel compression benchmark
using the zlib provider, running on various numbers of threads. Figure 5. Impact of
Multi-Threading on zlib Provider (SW Only, No Acceleration, No HT) shows the results
of compressing a 1 GB file using the zlib provider, in terms of both the elapsed time and
the CPU time.
Note: The data was generated using the Calgary corpus, a compression level of 1, a single
chunk per thread (by specifying a chunk size of 1 GB), and an affinity mask that caused
threads to be scheduled on only one logical core (lcore) per physical core (pcore),
meaning that hyperthreading was effectively disabled. See Section 8.0 Methodology for
details. The actual command line used was as follows:
parcomp -pzlib -icalgary.1G -otest1 -l1 -c1000000 –yffff –tn
where 1 ≤ n ≤ 16.
Programming Intel®

Programming Model
Figure 5. Impact of Multi-Threading on zlib Provider (SW Only, No Acceleration, No HT)
Key data points to note:
• The elapsed time decreases as more threads are added. For example, it reduces
from 19.18 seconds with a single thread to 1.41 seconds with 16 threads.
• The CPU time (see Section 8.5.4 Compression CPU Time) stays almost constant, as
the same CPU cycles are simply spread across multiple cores. The CPU time
increases very marginally as threads are added, probably due to the thread
management and context switching, and reassembly of the output.
• In all cases, the compression ratio (not shown) was constant.
The impact of enabling hyper-threading was also measured. This led to a reduction of
approximately 20% in the time taken to compress the data (e.g., running two software
threads on the two lcores of the same pcore led to a reduction in time of approximately
20% vs. a single software thread on one lcore).
Note: The reported CPU utilization in this case increased by approximately 50%. However,
this may be ignored since hyperthreading essentially uses otherwise unused cycles of
the physical core.
Now, again consider the offload case below.
Programming Intel®

Programming Model
6.2 Hardware, Synchronous, Single Threaded
The naive way to introduce acceleration is to offload each request from a single-
threaded application, using synchronous semantics (sometimes called blocking mode).
In this case, the application waits for each request to complete before sending the next
request. However, this does not make optimal use of the hardware (only one engine is
being used), nor does it make optimal use of the CPU cores (only one core is being
used, and that is idle for much of the time waiting for the response). This is illustrated in
Figure 6. Calling Hardware Synchronously from a Single Thread. The yellow bars
represent cycles spent doing the offload to hardware, e.g., creating descriptors, writing
to doorbells, handling responses, etc.
Figure 6. Calling Hardware Synchronously from a Single Thread
Note: The diagram illustrates the case where the job is broken up into multiple chunks, rather
than offloading a single larger job. There are several reasons why this is done, aside
from parallelism. For example, the application may wish to minimize the amount of
physically contiguous, pinned memory required; or be able to decrypt or decompress
subsets of the data independently, e.g., at a disk block boundary; or the hardware may
operate optimally at a given buffer size, as seen in the earlier Figure 3. Compression
Ratio vs. Chunk Size.
To see how this compares to the zlib-based software implementation, in terms of both
compression ratio, elapsed time, and CPU time, see Figure 7. Time Taken to Compress
1 GB using Software and Hardware.
Note: The data was generated using the Calgary corpus, and a compression level of 1. The
chunk size was chosen to be optimal for each case, i.e., for zlib, a single chunk (chunk
size 1 GB) was used, and for hardware, a chunk size of 64 KB was chosen. See Section
8.0 Methodology for details. The actual command lines used were as follows:
parcomp -pzlib -icalgary.1G -otest1 -l1 -c1000000 –j1 –t0
parcomp -pqat -icalgary.1G -otest1 -l1 -c64 -j1 -t0
Programming Intel®

Programming Model
Figure 7. Time Taken to Compress 1 GB using Software and Hardware
• Even using the hardware synchronously and single-threaded, it runs faster in this
“single-stream” mode than software (6.3 vs. 19.6 seconds), for essentially the same
compression ratio (38.5% vs. 38.1%).
• The CPU is mostly idle (waiting for the job to complete), so it consumes less power
than the case where it is 100% utilized implementing compression in software.
Alternatively, other processes can be scheduled to run on the core during this idle
time without impacting on the compression time.
By taking advantage of parallelism, optimizing further the utilization of both the
accelerator and the CPU can be done. There are at least two ways to take advantage of
parallelism: using multi-threading or using asynchronous (non-blocking) calling
semantics. Both methods can improve the elapsed time taken to compress the file, but
the asynchronous mechanism does so at a much lower cost in terms of CPU cycles.
Each model is reviewed below.
Programming Intel®

Programming Model
6.3 Hardware, Synchronous, Multi-Threading
As shown in Figure 8 (a), multi-threading can be used as done for the software case
above. These threads may run on separate cores, as shown here, or since the CPU
thread was under-utilized they may run on a single core [in which case it looks more
like Figure 8 (b)].
Figure 8. Calling Hardware from Multiple Threads or Asynchronously
Figure 9. Hardware Offload from Multiple Threads Synchronously shows the data for
various numbers of threads.
Note: The data was generated using the Calgary corpus, a compression level of 1, and a
chunk size of 64 KB. See Section 8.0 Methodology for details. The actual command line
used was as follows:
parcomp -pqat -icalgary.1G -otest1 -l1 -c64 -j1 –tn
where 1 ≤ n ≤ 16.
Programming Intel®

Programming Model
Figure 9. Hardware Offload from Multiple Threads Synchronously
• The elapsed time drops as the number of threads increases, as expected. For
example, using 16 threads, it drops from approximately 6.4 to 0.9 seconds.
• The total CPU utilization, however, increases due to the context switching and other
costs, from about 6.9 seconds (at one thread) to nearly 10 seconds (at 11 threads).
Note: The peak in CPU utilization for five threads, and the drop at three curves, are
unexplained, but are reproducible; this is an area for future investigation.
Programming Intel®

Programming Model
6.4 Hardware, Asynchronous
Alternatively, as shown in Figure 8 (b) above, the Intel®
QuickAssist Technology API can
be used asynchronously. As shown by the data in Figure 10. Calling Hardware
Asynchronously from a Single Thread, this tends to be more efficient than multi-
threading, in that no context switching between OS threads is required.
Figure 10. Calling Hardware Asynchronously from a Single Thread
• The elapsed time drops as the number of jobs outstanding increases. For example,
by having up to 24 jobs outstanding, it drops from approximately 6.3 to 0.73
seconds. With 16 jobs outstanding, the elapsed time is marginally less than using
16 threads.
• The CPU utilization also drops correspondingly, from about 6.8 seconds (at one
thread) to about 1.7 seconds (at 24 jobs). This is significantly less than the CPU time
using multi-threading. The likely explanation is that in the multi-threading case,
each thread spends lots of time polling for responses, with just a single response
expected. In the async. case, there is a single thread polling, and this is mostly when
there are multiple jobs outstanding, which is more efficient.
Note: The fact that both elapsed time and CPU time remain flat in going from 3-6 jobs
outstanding is unexplained, but is reproducible; this is an area for future investigation.
Programming Intel®

Programming Model
6.5 Multi-Core and Hyper Threading
It was noted earlier for the zlib provider (i.e., a software implementation of
compression), hyper-threading led to an improvement of in the order of 20% in
compression time.
For the hardware offload case, this was not measured since the asynchronous offload is
more efficient than multi-threading.
6.6 Multi-Socket and NUMA
As described in Section 8.0 Methodology, the system under test is a two-socket system
with a single Walnut Hill card. The impact of NUMA on the performance of the sample
application was not measured for this document; this is an area for future investigation.
6.7 Batch Submissions
Another mechanism to achieve multiple outstanding jobs is to submit multiple jobs in a
single request. An application which has multiple jobs to submit can submit them all in
a single call. This is called batch submission, and it is supported on the Intel®
QuickAssist Technology API (specifically the so-called “data plane” API for crypto and
compression).
Note: This is supported in asynchronous mode only.
A variation on this theme, also supported on the data plane API, is the ability to submit
a single job to the request ring but to defer updating the tail pointer (i.e., ringing the
“doorbell”) until after a number of requests have been submitted. The benefit of this is
that it allows the cost of the MMIO operation (ringing the doorbell) to be amortized
across multiple requests. This is also supported on the Intel®
QuickAssist Technology
data plane API. Intel has used this technique in an IPsec implementation.
The impact of these techniques was not measured for this document; this is an area for
future investigation.
6.8 Polling vs. Interrupts
As described earlier, when the hardware has completed a job, it writes the output data
to the DRAM, writes a response descriptor to ring memory (which is also in the DRAM),
and then writes to a doorbell register to indicate that the response is available.
Applications can decide whether to have this doorbell trigger an interrupt or whether
they wish to poll. A detailed comparison of the approaches is outside the scope of this
document; however, some of the key pros and cons of each approach are highlighted.
• Latency-sensitive applications may wish to enable interrupts, which can be
configured to fire as soon as a response is available, minimizing latency.
Programming Intel®

Programming Model
• Throughput-oriented applications may wish to use polling to avoid excessive
interrupts.
• Real-time systems tend to prefer polling, as interrupts can make it difficult to
guarantee meeting deadlines.
For applications that do use polling, there are two additional design decisions that need
to be made:
• Polling frequency: A higher polling frequency will reduce the latency but at the cost
of potentially higher CPU utilization. Ideally, an application should poll only when
there are known to be outstanding requests, and then at the lowest frequency while
maintaining acceptable latency. Some tuning may be required to identify the
optimal polling frequency for your application.
• Inline polling vs. separate polling thread: Inline polling (i.e., polling from the same
thread that submits jobs) will typically avoid context switching overhead.
Note: The parallel compression application described in this document uses polling
(specifically inline polling) at a high frequency. For this document, the impact of polling
vs. interrupts were not measured; this is an area for future investigation.
Programming Intel®

Programming Model
6.9 Programming Model Summary
The different programming models are compared side-by-side in Figure 11.
Figure 11. Comparison of Different Programming Models
• Using hardware, a reduction in elapsed time and CPU cycles was demonstrated to
compress a 1 GB file by more than 20x. This allows jobs requiring compression to
run in a shorter elapsed time, and frees up the CPU cores for value added
applications.
• An asynchronous programming model was implemented underneath a
synchronous API at the compression framework level. This means that existing
software applications can use this without having to change to an asynchronous
programming model.
Note: Implementing a synchronous API on top of an asynchronous implementation is an
example of the so-called “half-sync/half-async” design pattern as described in
reference [5].
The API defined for compression has simple stateless semantics, similar to the
Windows* RtlCompressBuffer API. The zlib API has more complex semantics. For
applications which specifically depend on behavioral compatibility with zlib, other
changes may be required. These are outside the scope of this document.
§
Programming Intel®

Memory Model
7.0 Memory Model
Software running in a user space application typically uses “virtual memory”, that is,
memory which is virtually addressed, virtually contiguous and paged. Hardware, on the
other hand, requires “physical memory”─memory which is physically addressed,
physically contiguous, and pinned (i.e., guaranteed to be resident in memory, at least
while the hardware is operating on it).
Note: This is because the hardware needs to be able to perform Direct Memory Access (DMA).
Future instantiations of the hardware may be able to perform DMA on virtual memory,
but for now, the above requirements apply.
Before sending a request to hardware to operate on data, that data needs to be either
copied to physical memory, or the pages need to be pinned and an SGL of physically
contiguous regions needs to be created. Both of these approaches add to the offload
cost.
In the sample application, the buffer copying approach was chosen. The time spent
performing buffer copying was measured. In the case of the Intel data, a total of
approximately 1.385 GB of data was copied (1 GB from the uncompressed buffer to the
pinned source buffers and approximately 385 MB of data from the pinned destination
buffers to the compressed buffer) in an average of about 0.2 to 0.25 seconds (or
approximately 0.4 to 0.5 cycles per byte, on a 2.7 GHz CPU core). Compared to the cost
of compression in software (which was measured at in the order of 40 cycles/byte), this
is relatively insignificant, although it may be more significant for smaller files or other
workloads.
§
Programming Intel®

Methodology
8.0 Methodology
Intel wanted to be able to compare the Intel®
QuickAssist Technology-based
implementation of compression to software implementations on various metrics
(compression ratio, time to compress a given file, CPU time, etc.). To that end, a simple
compression framework was created with the ability to plug in both hardware and
software implementations under its API. Then, a simple benchmark application was
created to call these implementations with various input parameters to measure the
impact of changing certain variables. The code is available from reference [6].
The software is illustrated in Figure 1. Parallel Compression Software Stack. A brief
description of the various layers is included below.
8.1 Benchmark Application
The benchmark consists of a program which reads a file from the disk into the DRAM
and compresses it using the compression framework. It then writes the output buffer to
the output file.
It takes various input parameters which allow us to measure the impact of varying
certain variables, as described in Section 8.4 Inputs.
It also measures various outputs and reports them to standard output, as described in
Section 8.5 Outputs.
8.2 Compression Framework
A simple compression “framework” is defined that consists of a simple “consumer API”
for an application, with the ability to register different implementations of that API,
called providers, via a separate “provider API”. Calls to perform compression are passed
on directly to the specified provider.
Note: An alternative would have been to use the existing zlib library as the framework, using a
zlib shim; that too would have allowed comparison of hardware and software
implementations. There were a few reasons why Intel chose not to do that for the
purposes of this document:
• The zlib software implementation and API have some unique behaviors/semantics
which add complexity to the zlib shim. The implementation for this document was
kept as clean and simple as possible.
• Intel wanted the flexibility to pass certain other parameters to the implementations
(chunk size, etc.) which were not easily accommodated on the zlib API.
Programming Intel®

Methodology
• Intel wanted to emphasize that the principles described herein apply not just to
zlib, but also to other compression implementations. In fact, many of the principles
apply equally to crypto libraries, as well.
• The zlib library does not provide a framework for adding providers
(implementations) under the hood; the zlib shim is therefore implemented as a
patch. For the purposes of this document, a clean separation was maintained
between the API and the implementations.
The providers above are called in the context of a software thread. The framework also
supports multi-threading. In this case, the framework will create the specified number
of threads (using the pthread library) and divide the input buffer into that number of
segments of approximately equal size3
. Each thread is then given one of the segments
to compress, in parallel. The main thread then waits for all the child threads to
complete, reassembling the output buffers as they complete, as shown in Figure 12.
Multi-Threading and Segments.
Figure 12. Multi-Threading and Segments
The API was chosen to map fairly closely to the APIs of several typical compression
libraries, specifically the zlib deflate() API and the Microsoft* Windows
RtlCompressBuffer() API (see reference [7]). The API is synchronous. However, it takes
an arbitrarily long input buffer, which allows implementations the freedom to compress
in parallel under the hood.
3
The last segment may be slightly larger than the others if the input file size was not an
integral multiple of the number of threads.
Programming Intel®

Methodology
The API is as follows:
// declare anonymous structure
struct CfProvider;
struct CfParams {
int level;
int deflateWindowSize;
int bStatic;
int chunkSize;
int maxJobsOutstanding;
int bDebug;
int numThreads;
int coreAffinity;
};
int cfGetProvider(
char* name,
struct CfProvider **ppProvider);
int cfCompressBuffer(
struct CfProvider *pProvider,
struct CfParams *pParams,
unsigned char* uncompressedBuffer,
int uncompressedBufferSize,
unsigned char* compressedBuffer,
int compressedBufferSize,
int* compressedDataSize);
int cfDecompressBuffer(
struct CfProvider *pProvider,
struct CfParams *pParams,
unsigned char* compressedBuffer,
int compressedBufferSize,
unsigned char* uncompressedBuffer,
int uncompressedBufferSize,
int* uncompressedDataSize);
8.2.1 Chunks
The compression provider (within a given thread) operates on chunks, the size of which
is an input parameter. Each chunk is compressed statelessly, yielding a deflate block.
These deflate blocks are concatenated in the output buffer. Each deflate block is
preceded by a header, which in this implementation (for simplicity) is simply a 4B
length indicating the length of the deflate block, as shown in Figure 13. Compressing
in Chunks.
Programming Intel®

Methodology
Figure 13. Compressing in Chunks
8.2.2 Buffer Sizes
When compressing a data buffer, the size of the compressed (output) buffer is not
known in advance. On most compression APIs, if the output buffer is not big enough to
contain the compressed data, then the function returns with a status indicating that the
compression is not complete, and the application is expected to provide additional
output buffer space and call the API again. Handling this so-called overflow condition,
while supporting both multi-threading and asynchronous APIs, would have added
significant complexity. Therefore, the above API makes the simplifying assumption that
the output buffer is big enough, otherwise it fails and exits. The sample benchmark
application allocates sufficiently large output buffers to ensure that this does not
happen for the input data used in the test cases (it currently assumes the compression
ratio will be in the range 20% to 200%). For different corpus data, with different
compression ratios, this may need to be modified.
Note: To see how a production-ready implementation might deal with overflows, see Section
8.3.2 qat Provider.
8.3 Providers
The framework defines the concept of a “provider” or implementation of the API. Two
providers of this API were developed and can be selected from the command line using
the –p (provider) flag, one based on zlib and one based on Intel®
QuickAssist
Technology.
Programming Intel®

Methodology
8.3.1 zlib Provider
The zlib provider uses zlib version 1.2.8.
The deflate API is called once for each chunkSize bytes, with the flush flag set to
Z_FINISH.
Note: This was a simplification, which comes with a small penalty in both performance and
compression ratio; it keeps decompression simpler, so that each deflate block can be
decompressed identically regardless of whether it was the last deflate block within a
thread in the multi-threaded case.
Intel also experimented with different values for the flush flag, but with chunk sizes in
the order of 64 KB and for the Calgary corpus which was tested, the impact on
performance and compression ratio was not significant.
8.3.2 qat Provider
The qat provider implements the CF API using Intel®
Specifically, it uses the “traditional” data compression API in an asynchronous fashion
in the Linux user space. In the compression direction, it operates as illustrated in
Figure 14. Flowchart for QAT Provider - Compression Function, and as described below.
The callback function which is invoked when a response is received is described in
Figure 15. Flowchart for QAT Provider - Compression Callback Function.
Figure 14. Flowchart for QAT Provider - Compression Function
Programming Intel®

Methodology
• The Init stage consists of the following steps:
− It first allocates a compression instance, which maps a set of request and
response rings into the address space of the process. (In the multi-threaded
case, a separate instance is mapped for each thread, so that each thread can
poll its corresponding response ring.) It also allocates memory for the
intermediate buffers to be used by the instance when performing dynamic
compression.
− It then creates a session object.
− It then allocates a set of up to maxJobsOutstanding pinned buffers for each of
the source and destination buffers, one for each outstanding request. (Again, in
the case of multi-threading, this is done per thread.)
• It then goes into a loop which does the following until the entire input buffer is
consumed:
− If the number of jobs outstanding is less than maxJobsOutstanding, it sends a
request to the accelerator to compress a buffer of size chunkSize, and
increments the number of jobs outstanding. The request flushes the
compressed data and creates a complete deflate block on a byte boundary.
− Otherwise, it polls for responses. If one or more responses are available, the
corresponding callback function is invoked, which writes the number of
produced to the output buffer preceded by the size of that deflate block. It then
decrements the number of jobs outstanding, and updates pointers into the
output buffer so that subsequent output data is written to the correct part of
the output buffer. After each poll for responses, the CPU is yielded to allow
other threads to run.
• Once the entire input buffer has been consumed, it polls for responses until there
are no more jobs outstanding.
Figure 15. Flowchart for QAT Provider - Compression Callback Function
Note: If maxJobsOutstanding is set to 1, this is equivalent to using the API synchronously;
the application sends one request, and then waits for this to complete before sending
Programming Intel®

Methodology
the next request. However, by setting maxJobsOutstanding to larger values, up to that
number of jobs can be in progress at any time, all from a single software thread.
As noted earlier, the current implementation makes the simplifying assumption that the
output buffer is big enough to store the output of the compression. A production-ready
implementation may want to deal with this in a more robust fashion, for example as
follows:
• Over-provision the output buffers to ensure overflow is an exception case. If it
happens, it will impact performance.
• If and when an overflow is received on the response to request number N, then it
can be dealt with as follows:
− Write the compressed data produced thus far from the pinned destination
buffer to the output buffer.
− Re-submit the request to process the remaining (unconsumed) uncompressed
data.
− As each of the next maxJobsOutstanding -1 responses arrives, the response
should be “held”, i.e., do not write the produced data to the output buffer−the
location to which to write will not be known until the earlier request has
completed and the size of that response is known.
− When the response to the overflow request arrives, write the produced data to
the output buffer, and then process each of the “held” responses.
8.4 Inputs
The input files were all based on the Calgary corpus, downloaded from reference [8].
This was uncompressed and untarred into a directory, and the individual files were then
concatenated into a single file (using cat * > calgary). The size of this file is ~3.25 MB
(3251493 bytes).
For the purposes of benchmarking, Intel wanted a file of ~1 GB in size, to allow the tests
to run for a sufficiently long time. Intel also wanted files with different compressibility.
Therefore, two 1 GB files created from the calgary file above:
• calgary.1G: For this, multiple copies of the calgary file were appended to one
another until the file was ~1 GB in size (i.e., repeated invocations of cat calgary >>
calgary.1G).
Note: Using this file achieved up to 11.7 Gbps of throughput and a compression ratio of
~38.5%.
• cal4K.1G: For this, the first 4 KB of the calgary file was taken (using head –c 4096
calgary > cal4K), and appended multiple copies of this until the output file was ~1
GB in size (i.e., repeated invocations of cat cal4K >> cal4K.1G).
Note: Using this file achieved up to 22.8 Gbps of throughput and a compression ratio of
~13.7%.
Programming Intel®

Methodology
Other inputs to vary include:
• The provider (zlib or qat), to allow comparison of SW and HW
• Chunk size
• Compression level
• Direction (compression vs. decompression)
• Maximum number of jobs outstanding (for qat provider only; a value of 1 means
synchronous, while larger values allow for greater degrees of asynchronous behavior)
• Deflate window size (for qat provider only)
• Static (vs. dynamic) compression (supported on qat provider only)
Note: Dynamic compression involves creating a dynamic Huffman tree; in hardware, this
requires sending the output of the compression engine through a separate translator
stage. This can improve the compression ratio, while increasing the latency of each job.
With sufficient jobs outstanding, however, the impact on throughput is minimal.
• Number of threads (for multi-threading)
• Affinity mask to specify the cores to which software threads should be affinitized.
Note: On the system under test, there were two CPU sockets, each with eight physical cores
and with Intel®
Hyper-Threading Technology/SMT enabled, for a total of 32 logical
cores. Some examples of the affinity mask required to identify, for example, thread 0 of
all physical cores on socket 0, can be seen in Table 5. Affinity Mask.
Table 5. Affinity Mask
Affinity Mask (–y) Socket 0 Socket 1
Thread 0 Thread 1 Thread 0 Thread 1
All 8 Physical
Cores
0x000000ff 0x00ff0000 0x0000ff00 0xff000000
Programming Intel®

Methodology
The full set of command line parameters are described by invoking parcomp with the –
h flag:
-p providerName Specifies the provider (implementation). Should be
one of qat (default) or zlib.
-c chunkSizeInKB Chunk size, in KB. Default is 64.
Files will be divided into chunks of this size
(the last chunk may be smaller), and each chunk
is compressed separately. State is not maintained
between chunks.
-l compressionLevel Compression level. Default is 1.
-d Decompress the input file. Default is to compress.
-v (or -g) Verbose (or debug)
-x numLines Print a summary of the inputs and outputs in a
comma-separated variable (CSV) format for easy
importing into a spreadsheet. Specify numLines as 1
for data only, or 2 to also include a header
summary
-t numThreads Creates specified number of threads, splits input
file into numThreads (near-)equal chunks, and
performs the operation in each thread. The maximum
number of threads is 32.
-f cpuFreqInMHz Specifies the CPU frequency in MHz. If not
specified, this will be measured (takes approx. 1
second)
-n numIterations Specifies the number of iterations (allows you to
run same compression N times. Default is 1.
-y coreAffinityInHex Specifies the cores to which threads should be
affinitized, as a bitmask, using hexadecimal
notation, e.g. -af means the first four cores.
Default value is 0xFF, which on a platform with
2-sockets x 8 pcores/socket x 2 lcores/pcore
affinitizes to socket 0, pcores 0-8, lcore 0
-a coreAffinityInDec Specifies the affinity in decimal; see -y for more
info
-h Print this help message
The following options are applicable for the qat provider only:
-j maxOutstandingJobs Maximum number of outstanding jobs (requests) that
may be outstanding at any one time. Default is 1.
The maximum number of jobs outstanding is 256.
-w windowSize Deflate window size. Default is 5.
-s Static compression. Default is dynamic.
8.5 Outputs
For each run of the benchmark application, the following was measured.
8.5.1 Compression Ratio
This is computed as the compressed file size divided by the uncompressed file size.
8.5.2 Compression Elapsed Time
The time taken to compress the buffer is computed by taking a timestamp (using
rdtscp) before and after the compression. The difference in cycles is translated into
Programming Intel®

Methodology
seconds by dividing by the CPU clock frequency. This was measured on the platform
used for testing as being ~2694 MHz.
Note: The time was measured at the cfCompressBuffer() API.
It does not take account of time taken for the following:
• Allocation/freeing of DRAM memory buffers to hold the input and output files
• Reading/writing of the input/output files into/out of DRAM
• Parsing command line parameters
It does take account of time taken for the following:
• For multi-threaded cases, creation/deletion/synchronization of threads
• For hardware (qat provider):
− Allocation of instances (i.e., mapping rings to user space) and sessions
− Allocation/freeing of pinned buffers for data
− Copying of the data, one chunk at a time, between the input/output buffers and
the pinned buffers passed to hardware
• For software (zlib provider):
− Stream initialization (deflateInit()) and termination (deflateEnd())
Note: As noted below, the throughput can be derived from the elapsed time. For hardware,
the focus is on its throughput. However, the graphs in this document typically chart the
elapsed time (seconds per GB) rather than the throughput (Gbps), because in many
cases, the hardware is not being fully utilized (e.g., when the number of jobs
outstanding is less than about 18). Focusing on the throughput in these cases would be
misleading.
8.5.3 Throughput
Throughput, measured in Mbps, is computed as the uncompressed file size divided by
the elapsed time taken to compress the file, as described above.
8.5.4 Compression CPU Time
This is the total CPU time measured across all cores in a multi-core system. It was
measured by reading from /proc/<pid>/stat just before and just after calling the
cfCompressBuffer API, and extracting the user and system times from that. (This was
correlated with the output of top to validate the correctness of the methodology.)
Note: CPU time is measured rather than CPU utilization because it takes into account the
number of cores on which it applied, and the duration. For example, a value of 1 second
can be made up in infinitely many ways, including:
• 100% of one core for 1 second
• 100% of two cores for 0.5 seconds
Programming Intel®

Methodology
• 50% of two cores for 1 second
8.6 Platform Description
The platform on which the benchmark was performed was as follows:
• Hardware configuration
− Platform: Crown Pass (two CPU sockets)
− CPU: Each CPU is an 8-core Intel®
Xeon®
CPU E5-2680 (formerly codenamed
Sandy Bridge) clocked @ 2.70 GHz
Note: There are a total of 32 hardware threads (2 sockets x 8 cores/socket x 2 HW
threads/core)
− DRAM: 32GB of DDR-3 @ 1333 MHz
− Accelerator: Intel®
QuickAssist Adapter 8950, featuring the Intel®
Communication Chipset 8955 controller (formerly codenamed Coleto Creek),
SKU 2
• Software Configuration
− BIOS: Disabled speed step (EIST) and turbo
− OS: Fedora* Core 16 (3.6.11-4.fc16.x86_64)
− Intel®
QuickAssist Technology Driver/Firmware Version: 2.1.0
§
Programming Intel®

Related Work
9.0 Related Work
In addition to the recommendations in this document, also consult this reference [9],
which describes some additional key design decisions that should be considered, and
other techniques that may be employed, in order to achieve optimal performance when
integrating applications with Intel®
§
Programming Intel®

Future Work
10.0 Future Work
Future work may include the following:
• Measure the performance impact of using buffer copy vs. dynamic pinning of
pages.
• Measure the performance impact of polling vs. interrupts.
• Measure the performance impact of polling from the main thread (as is done in the
example used in this document) vs. polling from a separate thread (as is done by
much of the sample code shipped with the driver).
• Measure the performance impact of NUMA.
• Measure the performance impact of batch submission.
• Understand some of the anomalies in the data mentioned earlier in this document.
• Develop similar sample benchmarks for public key crypto and symmetric crypto,
illustrating the principles from this document.
§
Programming Intel®

Summary of Recommendations
11.0 Summary of Recommendations
As demonstrated by the measurements taken using the parallel compression
application developed as part of this document, the following recommendations
should be followed in order to achieve optimal performance when programming Intel®
QuickAssist Technology based accelerators:
• Identify opportunities for parallelism in your application, as described in Section
5.0 Parallelizability.
• To take advantage of this parallelism, ensure to keep the hardware busy by having
sufficient jobs outstanding.
− Ideally, use an asynchronous API for accessing Intel®
QuickAssist Technology,
since this is the most efficient way to have multiple jobs outstanding.
− As an alternative, use multi-threading.
• To minimize the impact to the application, consider “hiding” the parallelization
“under the hood” of a synchronous API, as demonstrated in the parallel
compression example used in this document.
• Aim for NUMA locality as far as possible. Design the application so that data stored
in memory associated with a given socket is processed by software running on
cores on that socket, and by accelerators which are physically connected to that
socket.
§
Programming Intel®

Programming Intel® QuickAssist Technology Hardware Accelerators for Optimal Performance

Recommended

More Related Content

What's hot (20)

Similar to Programming Intel® QuickAssist Technology Hardware Accelerators for Optimal Performance (20)

More from DESMOND YUEN (20)

Recently uploaded (20)

Programming Intel® QuickAssist Technology Hardware Accelerators for Optimal Performance