Generic SAN Acceleration White Paper DRAFT
- 1. * DRAFT * Applying Highly Optimized SSD Performance to the Application Server Environment By Michael S. Mendola Sr. Storage Architect A Mendola White Paper
- 2. Contents Introduction 2 Is SSD Something New? 2 The Spin on How SSD Stores and Retrieves Data 3 The Three Keys: Form Factor, Location and Integration 4 Making Globally Accessible SSD Truly Global 6 Summary 7 Introduction SSD - Solid-State Disk. It’s probably a safe bet that by now you are at least aware that the major alternative to storing data on spinning disk is replacing that disk with random access memory which acts like spinning disk but is much higher performing. Have you already considered augmenting or even replacing spinning disk with SSD? Maybe you’ve just begun gathering information or are even ready to start infusing SSD into your environment. In either case, we know that you will benefit greatly from the information provided here. Is SSD Something New? You might be surprised to learn that the use of memory-based data storage actually preceded magnetic media-based data storage. Called Auxiliary Memory Units non-semiconductor memory, made persistent using electronic capacitors, was the sole data storage medium until spinning magnetic drum and disk systems were designed that were reliable and inexpensive enough to replace AMU storage. In the 1970’s, SSD made a come-back using semiconductor memory which far eclipsed disk storage in performance but was still prohibitively expensive to use in any meaningful quantity. Termed RAM disk and semiconductor disk, these devices were housed in large frames and used battery backup to retain active data. They were used mainly in supercomputers but eventually made their way into the mainframe strata. Here, they were used mainly for system OS and virtual storage page pool use. Soon, due to performance, reliability and price-point improvements, high performance memory was becoming a
- 3. viable alternative to spinning disk for smaller systems and even found its way into the burgeoning field of personal computers (known as RAM disk). Even as memory is now becoming primary “disk”, it has been in use for many years to augment the performance of spinning disk. Memory almost always used as the front-end data cache in storage array controllers and is even found in the electronics of the discrete disk drive themselves. Today, with the advent of inexpensive non-volatile flash-based memory which provides “meaningful” data storage capacities and the elimination of the need for batteries to retain data, memory- based disk systems (now called SSD) have made their way into the mainstream. They are available to vastly increase performance across all compute platforms and applications. SSD is not only here to stay, but as it evolves promises to eventually replace spinning disk for almost all data storage requirements, with the probable exception of those which require protracted retention periods or a final resting place in static archives. Finally, memory systems now used in SSD devices are being specifically engineered and tuned to provide unprecedented I/O performance (measured in IOPS), reliability, longevity and energy efficiency. Replacing spinning disk forever is indeed the goal, and the race is just beginning! The Spin on How SSD Stores and Retrieves Data Before moving into examining how SSD is best used, we will first briefly review how it operates. This is important to understand as we set the stage for how to best inject SSD as an accelerator in the primary storage environment. Two types of memory are used in SSD devices. These are non-volatile flash memory and the much faster, but volatile (and much more expensive) DRAM memory. Without delving too much into the technical aspects of each, suffice it to say that flash memory is most widely used because of it being more cost effective and its ability to hold data even when power is removed. DRAM memory has neither of these qualities, but does perform at higher speeds with lower latency and is much more efficient in the way it stores and retrieves data. Flash memory is made of a special type of read-only memory called EEPROM (Electrically Erasable Read-Only Memory) that can be erased and “reprogrammed”. EEPROM “read-only” memory is best suited for use in semi-permanent programming storage devices where data (typically firmware) is stored once and read many, many times. When used in SSD devices, flash memory is expected to both retrieve and store data continuously. Not specifically designed for this type of heavy store-erase cycle intensity, the memory cells wear out, eventually becoming useless (typical flash write/erase life span is about 100,000 cycles). Also, these devices actually become slower as they are filled with data. As a counter measure to the write/erase life span problem, almost all SSD designers build-in wear leveling techniques which count writes and dynamically re-maps data blocks to other
- 4. areas, spreading the wear over a broader cross section of the SSD device. Another technique is to configure the SSD device with more storage capacity than is actually presented for use. This, in conjunction with wear leveling, work well to bring flash-based SSD MTBF numbers more in line with that of their spinning SATA disk counterparts. To increase speed and reduce latency even further, designers will configure simultaneous access across multiple flash memory devices. DRAM-based SSD devices do not have the write/erase life span shortcoming, but must always have a constant source of power to continually refresh data on the device. The fastest DRAM memory uses data refresh cycles which are closer to system CPU speeds. Still prohibitively expensive to use as the sole memory in SSD devices, DRAM is more often used as the front-end cache to flash-based devices in some of the more robust enterprise-class SSD memory appliance systems (more on this later). The Three Keys: Form Factor, Location and Integration Let’s look at form factor. The most well known and immediately usable SSD form factor is that which is configured to directly replace 5.25” or 2.5” spinning disk devices. Here, the SSD device is packaged exactly like the spinning disk, even using the same SATA or SAS interface. Most SSD units are purchased in this form because it is readily installed and immediately usable. SSD “drive” in standard 2.5” form factor with SATA interface Using another familiar form factor are those devices packaged as discrete PCI cards. Geared toward low-end servers, these devices plug directly into the computer PCI bus and have the advantages of greater speed and memory density in a smaller footprint. This form factor is best used to accelerate database system transaction log and index performance. The key issue is ensuring driver availability and OS compatibility. SSD “drives” in I/O card form factor using PCI-Express interface The last major form factor is the memory appliance. Here, the memory is configured as discrete DIMM’s (Dual Inline Memory Module) in a 2U “memory server” appliance. SSD configured in this form factor is by far the highest performing and most flexible implementation available. Intended for serving disk-emulated storage to high-end servers and server clusters,
- 5. memory appliances are specially designed with advanced scalability, serviceability and maintenance features. Specialized DIMM’s provide the optimum in speed and longevity. Failed DIMM’s in a memory appliance may even be dynamically replaced by hot spares and failed units replaced during normal operations. Opened memory appliance showing memory DIMM’s. Appliance uses two PCI-Express interface cards to connect to servers (not shown). Now that you have a better feel for how SSD is packaged for use, it’s time to address the most important criteria. You’ve heard it said in the real estate market that the most important factors are location, location and finally, location. The same holds true when determining where best to inject SSD. Choosing the right form factor is the first step as this determines where SSD can be placed in the data storage environment. Because of its high performance-to- expense ratio, there are many factors to consider when planning to place SSD into the storage environment. The most important step is to understand as clearly as possible both the current and anticipated primary storage capacity and performance needs for the most critical business applications. For all but the smallest of organizations this is no trivial task and we suggest brining in a trusted advisor or consultant which is not only a subject matter expert, but vendor- neutral. Being vendor-neutral is a key attribute for any hired help because, as you might imagine, SSD form factor and placement options are tightly constrained by storage system hardware and software and storage services architecture and configuration capabilities – it can’t be expected that any storage vendor or their resellers will suggest injecting SSD into your environment in ways that their systems cannot use it. Finally, let’s take a look at SSD integration into the application environment. At first blush, simply replacing spinning disk units with their SSD counterparts seems like the most simple and efficient way to bring SSD into the environment. This assumption may be correct, but, like many other things in life, it depends on at least a few other considerations. Simply replacing spinning disk with SSD can overwhelm the server or storage array in the SAN, especially if too many are installed. This is because server- internal and even SAN storage array controllers are designed to work efficiently within the limitations of spinning disk; SSD performance will be hampered and you won’t achieve the performance you’ve paid for. Just replacing spinning disk inside a server (using either disk- or PCI card-based form factors) will improve the application server’s disk I/O performance, but can break the server SAN-to-disk array relationship that yields the many benefits of networked storage in a SAN environment, and t his includes breaking the shared storage requirements of clustered application servers. Also, placing discrete SSD devices into a server or even one or more SAN storage array limits the benefit to only those applications which can have access to these
- 6. devices by connection. Further, discrete SSD devices typically must become part of a RAID group where portions of the device are used to store data along with portions of other devices, whether these are spinning disk or additional SSD devices in that particular RAID group. This can be a problem because you must buy more SSD devices (and further loose capacity to RAID level and formatting) and performance is limited to the lowest performing devices in that group. SSD truly is best implemented where it can be applied in a global way to spread the benefits as widely as possible across many/all enterprise applications. SSD should placed as a globally available resource in a shared storage environment such a in a SAN-based array (due to controller and other bottlenecks as discussed above, your choice of array can make or break your SSD ROI. Work with your consultant in choosing from newer array architectures which are specifically designed to use SSD as a globally accessible resource without limiting its performance). The ultimate way of making SSD a globally accessible resource while even amplifying its performance is using the memory appliance-based form factor and attaching this to a specialized storage controller system (as opposed to disk storage controllers within storage arrays). These specialized storage controllers are designed to operate independently of the disk devices they control (and the protocols though which they’re accessed) by vitalizing this physical disk capacity. SSD is positioned as additional “disk” under these specialized controllers which make this a “layer” of targeted application acceleration across the entire field of application servers. (You be have guessed that storage virtualization is a very deep topic in and of itself, and as such is beyond the scope of this white paper. Your consultant should have a broad and detailed understanding of this subject and should be able to help you understand the many ways storage virtualization can be architected in greatly benefit your current and future storage environment efficiency in use and buying power in many ways). Making Globally Accessible SSD Truly Global You’ve chosen the form factor, location and integration points for your ultra-high performing and SSD, and you’ve paid a good chuck of change to buy a relatively small amout. You know that you are certainly on the right track, but if you haven’t already, you will soon realize that you can’t afford enough of it to make all of your application performance wishes come true. The remedy is teiring your SSD, using it as a “layer” in the primary storage “stack”. Storage teiring matches varying storage resources to specific application performance and capacity requirements. This technique achieves the best balance of cost vs. performance. Tiering is a good method of assigning storage based specifically on efficiency and performance criteria. This method allows you to place data only on the storage that most closely meets the real-world I/O performance demand. In fact you are
- 7. probably using tiered storage architecture already, placing most of your user data on SATA disk while keeping your most demanding database and log data on SCSI SAS or Fibre Channel disk resources. It is a simple matter to extend this idea by adding a final top-level tier of SSD capacity for none but application data of the most demanding nature. However, this standard tiering model, which works very nicely, now breaks down at the highest level when such a relatively scarce and expensive resource as SSD is used made the top tier in the I/O performance vs. cost stratification. Standard or static tiering practices are not optimal- especially where SSD is used - because although the appropriate storage is made to host the appropriate data, data I/O performance demands ebb and flow and the only way to accommodate this is to manually monitor and move data between teirs as necessary. An automated and intelligent data migration mechanism is required to make tiering really pay-off – especially when placing SSD into the mix. Locate and integrate SSD into a storage controller environment that will profile data I/O read/write frequency and other use patterns and use this information to automatically and interactively move data between tiers. Automatic tiering is important to make the best use of tierd spinning disk – it is imperative to achieving optimal global use of your most expensive and scarcest of all resources, the top-level SSD tier. Summary The use of memory as disk storage is not new. Improvements in technology, packaging and integration capabilities have made what is now called SSD a mainstream application accelerator available to individual users and organizations of every size and budget. Although at first glance using SSD seems relatively simple, bringing the benefits into all but the smallest compute environments requires understanding and experience with this technology and the storage environments through which it will be put to use. It is recommended that anyone considering bringing SSD application acceleration into their environment consider talking with an experienced consultant in the data storage field.