Buying Guide: Flash Arrays
At the recent EMC World, EMC made some bold claims about how its flash arrays were faster, better and more reliable than those of the competition. Let’s review EMC’s statements and flash array offerings, and also hear from some of its competitors about those claims, as well as what they offer as an alternative.
XtremIO is arranged in 10 or 20 TB X-Bricks, which are said to provide over 250 TB of effective capacity in one cluster and deliver over a million random IOPS. EMC makes a big deal of its “transformational architecture,” which was designed from the bottom up to take advantage of flash. It uses Intel processors in combination with software features such as content-based data placement and inline deduplication, as well as a dual-stage metadata engine and data protection.
“XtremIO never commits a duplicate block to flash,” said David Goulden, CEO of EMC's information infrastructure (EMCii) organization. “Deduplication is done inline.”
Jeremy Burton, president, products and marketing at EMC, said XtremIO is now the top product by revenue in flash storage. Without naming who the specific competitor was, he showed some graphs claiming that other arrays slowed to a crawl under certain conditions as they used a post-processing form of data deduplication.
The basic scenario was that this other array became very busy and so therefore backlogged deduplication, which then impacted performance and indeed capacity. To make his point, he offered a $1 million guarantee on flash performance for XtremIO and a trade-in program for other arrays.
OK, that’s the EMC view. Now what do competitors have to say about it?
PureStorage provides three basic units of all flash storage in its FlashArray family. The largest one, the FlashArray 450, supplies up to 200,000 IOPS, an average latency of less than 1 ms, 7 GB/s bandwidth and up to 250 TB of effective capacity (up to 70 TB of raw capacity).
The company is clearly at odds with the above EMC statements, having released one blog contesting EMC “propaganda” and another suggesting ways to obtain the one million from EMC. Jim Sangster, director, product & solutions marketing at Pure Storage, said FlashArray offers five forms of data reduction – four of which are inline. He attacked the EMC performance benchmark where they claimed 1 million IOPS as not reflective of a real-world scenario.
“Pure Storage delivers variable IO size, as opposed to a fixed IO size (i.e. 4K) providing better performance where it counts: in IO ranges of ~20K-60K,” said Sangster. “Architectures with fixed IO size like XtremIO have to break larger IOs up into their fixed size and the number of IOPS drops accordingly: 1M 4K IOPS drops to 1/2 M IOPS at 8K, and drops to ¼M IOPS at 16K, and so forth. With Pure’s variable geometry these can all be handled in 1 IO and provide higher numbers in the real world zone of IOPS use.”
He ended with a plug for FlashArray’s resiliency, its ease of adding/removing drives without impacting performance, and its snapshot capabilities.
“EMC’s new snapshots cause performance issues,” said Sangster. “Removing snapshots from the middle of a snapshot tree will cause performance degradation.”
EMC’s remarks about the horrors of post-processing deduplication seem to imply that this was a general problem affecting many flash arrays. But whatever competitor’s product it cited, none of those interviewed for this story were reliant on this approach. In addition to Pure Storage, Kaminario also supported inline deduplication and inline compression.
“There is no need to save deduplication for quiet times or run into issues with production performance,” said Ritu Jyoti, chief product officer at Kaminario. “The secret is in the employment of appropriate deduplication algorithm that consumes less CPU and has less metadata overhead.”
On top of 100-percent inline deduplication processing, he said that Kaminario K2 offers consistent performance for single and/or mixed workload environments. Like Sangster, she targeted XtremIO’s fixed 4 KB block size as an area of concern. This method, she said, results in limited bandwidth for the application and multiple IOs for each IO that is bigger than that fixed block size.
Kaminario boasts that its technology can generate two million IOPS with a bandwidth of 30GB/s in a single system. Latency is states as 120 microseconds on writes and 280 microseconds on reads.
Fusion-io COO Lance Smith is similarly on the offensive. His view is that to overcome the inherent slowness of hard drives, array vendors are forced to leverage dedupe to eliminate the amount of data written, which gives them a chance to keep up with the storage demands of an application’s workload.
“Fusion-io’s flash-based architectures are fast at both read and write transactions and do not need data processing overhead/elimination to deliver maximum storage performance for applications,” said Smith. “This is due to our low latency, high IOP, high bandwidth and symmetrical NAND flash based ioMemory storage technology used in our Fusion ION Accelerator all-flash appliance and our native PCIe server-side flash product lines.”
The company lists 1.7 Million random IOPS, 56 microsecond access latency and 23GB/s bandwidth for transactional and sequential workload acceleration. Additionally, incremental scalable storage of up to 32TB ioMemory capacity is said to be available to in 1U to 4U of space.
Flash Array Summary
There are further flash array vendors that Enterprise Storage Forum could interview on this topic and each might offer valuable insight. But what appears to emerge is the same old lesson – whatever the vendor says, there is no substitute for testing technology in your own environment and under your own specific workloads.
“In an era of consolidation and virtualization and the need to do more with less, it is increasingly important to deploy an all-flash array with the right architecture, one which offers the best fit at lower cost – and one that can scale as the customer needs,” said Kaminario’s Jyoti.
Photo courtesy of Shutterstock.