BACKGROUND IMAGE: iSTOCK/GETTY IMAGES
To get the right all-flash array (AFA) for your organization's needs, it is important to compare features of many of the leading vendors in the all-flash array market against specific criteria to help you get past the marketing hype. The data used in this article comes from manufacturer data sheets available on the respective public websites as of May 4, 2015. Our criteria for inclusion in this article is that the arrays were specifically built for flash storage and do not support hard disk drives.
At Demartek, we normally like to run performance tests, ease-of-use tests and other similar tests requiring hands-on experience with these storage systems, but this article is based on public information provided by the manufacturers. You can expect to see hands-on reports in the Demartek SSD Zone for several of these storage systems running real-world applications.
Most product vendors are in some stage of development of the next version of their product, or perhaps an entirely new product. It is quite likely that by the time you read this article, one or more product vendors will have "leap-frogged" ahead of some of their competitors for some of the features listed here. This is quite normal for the all-flash array market, so use this feature list as a guide, but not as the final arbiter or decision maker.
All about the data
As noted in the second article of this series, storage capacity is an important technical feature, but storage capacity is not what it used to be many years ago due to compression and deduplication. This affects the price you pay for an array, so it is important to review a few things about how all-flash arrays store your data.
First here are some definitions, taken from the SNIA dictionary. It is important to understand these three definitions, as you will see some or all of them used in the marketing materials provided by the manufacturers of all-flash arrays:
- Raw capacity: The sum total amount of addressable capacity of the storage devices in a storage system.
- Usable capacity: Synonym for formatted capacity, or the total amount of bytes available to be written after a system or device has been formatted ... less than or equal to raw capacity.
- Effective capacity: The amount of data stored on a storage system, plus the amount of unused formatted capacity in that system. There is no way to precisely predict the effective capacity of an unloaded system. This measure is normally used on systems employing data reduction technologies.
Data reduction technologies are much more feasible to implement in high-performance all-flash environments than they were in legacy hard disk drive environments, which is why these are becoming "table stakes" for AFAs.
Flash is fast enough to perform data reduction inline and because deduplicated or compressed data results in fewer writes to the flash media, it also increases endurance. Some of the AFAs implement data reduction technologies as a first pass inline and then follow up with additional processing behind the scenes with the data at rest. Vendors in the all-flash array market sometimes argue that the order in which they perform data deduplication and compression, along with their specific algorithm, makes their array better than other products.
Features such as thin provisioning, data deduplication, compression and some snapshot technologies fit under the general description of data reduction, data optimization, capacity optimization, space optimization or space efficiency. Most of the AFAs in this list support one or more data reduction technologies. Several of them require these features to be always on and do not allow them to be disabled. The particular combination of data reduction features and the algorithms used are part of each product's "secret sauce," making comparisons between them somewhat difficult. These combinations of data reduction technologies directly affect the effective capacity of a storage system.
A handful of vendors in the all-flash array market quote a fairly conservative data reduction ratio of 2:1 for storing data. Other vendors quote higher ratios of 10:1 or more. I often see a data reduction ratio of 5:1 or 6:1 advertised for mixed workloads for systems that implement data deduplication and compression. The advertised data reduction ratios should be considered guidelines for mixed overall data types, and generally not a guarantee, because different data types have different data reduction characteristics. Data deduplication works well for workloads such as virtual desktops or servers that use clones. Compression usually works well for database workloads. AFAs that have both data deduplication and compression can handle a variety of mixed workloads.
Data reduction ratios are used to compute not only the effective capacity of the storage system, but also directly affect the price per gigabyte of the system, as the price quoted is frequently based on the effective capacity. It would not be unusual to see prices based on effective capacity to be $2 per gigabyte or less for the large-scale capacities of these products. These prices are expected to continue to drop for the foreseeable future.
Many all-flash array product vendors provide performance ratings in terms of input/output per second (IOPS), throughput and latency in their marketing materials, but they don't all use the same criteria for the numbers they show. Because of this discrepancy, I am not including their published performance ratings in this product list.
However, to help you understand some of the data, here is some background information for storage performance. Typically, IOPS are shown for transactional (random) workloads, as these workloads tend to use smaller I/O block sizes. The throughput (or bandwidth) numbers are usually shown for sequential workloads with larger block sizes. If the AFA vendor includes IOPS or throughput numbers on its product datasheet, it will typically be for one block size for IOPS and a larger block size for throughput. They don't all choose the same block size to report, however.
Real-world workloads often use multiple block sizes together or odd block sizes, so benchmark data that isolates a single block size has some usefulness but does not tell the whole story. The following graphs highlight the typical results (on a relative scale) that I see for IOPS and throughput when we run storage performance tests in our lab, showing block sizes ranging from 512 bytes to 1 megabyte.
Latency, sometimes known as response time, can be a little trickier, but is becoming increasingly important for all-flash arrays. Because all flash arrays have much lower latency than storage systems based on hard disk drives, average latency of less than 1 millisecond (ms) is usually seen as the basic target.
With AFAs running transactional workloads, I typically see consistent sub-millisecond latencies within a reasonably tight range. However, the same all-flash array running an intense data warehousing workload will have several spikes in latency well above 1 ms and have a broader range of latencies over the period of the run. Therefore, the average latency highly depends on the workload or mix of workloads.
Because of this workload dependency, some vendors in the all-flash array market quote the minimum possible latency for their system regardless of workload, sometimes in the neighborhood of 100 or 200 microseconds (0.1 or 0.2 ms). Others simply quote an overall average latency, while others quote their best IOPS number (transactional workload) while maintaining less than 1 ms latency. Again, because of the wide variance in the criteria for publishing latency data, I have omitted latency from the list.
Because of the way warranties are handled, I need to provide an additional comment or two beyond the data in the table. The all-flash array manufacturers quote an initial warranty period such as one year or three years on their datasheets. Almost all of them offer an extension to the warranty period for an additional charge for a few more years. This additional charge needs to be considered into the total price.
The other criteria that I mentioned in the article "How to buy an all-flash storage array" are a little more straightforward, so I have simply included the data from the manufacturers into the table below. In some cases, I have listed a single model. In cases where the features were similar, I listed multiple models.
Determine if all-flash arrays are right for your organization
Review a guide on enterprise all-flash storage arrays