How bad is SSD performance degradation?

December 7, 2010 -- “It's a fairly well known fact that solid-state disk (SSD) performance can suffer over time. This was quite common in early SSDs, but newer controllers have helped reduce this problem through a variety of techniques. In part one of this two-part look at SSDs, we examine the origins of the performance problems and some potential solutions.”

That’s how Enterprise Storage Forum contributor Jeffrey Layton begins his exhaustive two-part series on SSD performance degradation.

For years (decades, actually), the focus on SSDs was on the exorbitant prices of the devices. Then the attention shifted to reliability, or endurance, issues. But SSD and controller manufacturers made great strides in those areas over the past couple years.

Now, the focus may be turning to performance degradation over time.

Layton’s articles are the best I’ve read on this subject. However, a warning: The articles are very long, very technical, and very detailed. But if you (a) are using, or considering using, SSDs (b) have sufficient technical credentials and (c) have a lot of time, I strongly recommend reading the articles.

Part 1 examines the issues that cause performance degradation in SSDs, and looks at some of the solutions (or “workarounds,” because all SSD solutions seem to involve trade-offs) that vendors have implemented.

Part 2 looks at technologies/issues such as write amplification, over-provisioning and the TRIM command, and then delves into some very in-depth testing of Intel’s X25-E SSD in ‘before’ and ‘after’ stress test scenarios. Check it out on Enterprise Storage Forum:

Fixing SSD Performance Degradation, Part 1

Fixing SSD Performance Degradation, Part 2

Startup’s SSDs top 1,000,000 IOPS

March 17, 2010 – According to independent test results, Pliant Technology’s Lightning Enterprise Flash Drives (EFDs), which are based on solid-state disk (SSD) technology, topped one million IOPS in a configuration with 16 drives.

Many storage vendors have issued press releases claiming performance of more than, or at least very close to, one million IOPS. Examples include Broadcom (press release), Emulex (press release), and Microsoft and Intel (blog post and video).

The reaction from most people (including me) would be “So what?” and/or “Who the heck needs one million I/Os per second?” It’s a legitimate response, because almost nobody needs that level of horsepower, even in the context of rapidly growing virtual server environments, multi-core processors, etc.

But that was also the response when 10GbE, 8Gbps Fibre Channel, 6Gbps SAS, etc. came onto the scene. And we know how those technologies are doing. The fact is, end users’ need for speed is insatiable, and if you don’t need it today – you will tomorrow.

The testing of Pliant’s SSDs was done by OakGate Technology, which supplies testing tools for IT technology providers.

As with any benchmark tests, it’s important to (a) have a grain of salt on hand and (b) look at the test configuration. The 1.1 million IOPS results were achieved with 16 drives and an 80/20 read/write ratio with 4KB block sizes.

The testing also revealed that that configuration can produce 2.3 million IOPS with a 100% read stream and 512-byte blocks.

Two ZT Systems Linux platforms with dual quad-core Xeon processors were used in the Lightning EFD tests, each running the OakGate FireOak test suite, which was controlled by a Java-based GUI running on Windows XP. Each system included four Adaptec 1045 SAS HBAs, with each HBA connected to two 3.5-inch, 300GB Lightning LS300 EFDs from Pliant.

For complete test details, see OakGate CEO and founder Bob Weisickle’s videos on YouTube:

· Benchmark 1: Real-world Workload
· Benchmark 2: Maximum I/O Rate
· Benchmark 3: Maximum Bandwidth
· Benchmarks 1-3 (complete version)

WhipTail: Software solves MLC SSD issues

July 29, 2009 – One of the more promising developments in the solid-state disk (SSD) drive space is the potential use of low-cost multi-level cell (MLC) NAND flash memory in enterprise applications and arrays, vs. the high-cost – but more reliable and durable – single-level cell (SLC) technology, as I mentioned in my previous post (see “Intel slashes SSD prices” ). This can basically be accomplished in two ways: via software or via controller enhancements.

Relative newcomer WhipTail Technologies is an example of a vendor that’s using software techniques to overcome some of the inherent limitations of MLC flash memory; namely, write amplification issues that limit the ability of NAND to perform random writes in an effective manner (a performance issue), and wear-out issues (a reliability, or durability, or endurance problem).

To address the performance part of the equation, WhipTail uses buffering (not caching) techniques, in which writes are aggregated into a buffer that’s sized to the erase block of the NAND, according to WhipTail CTO James Candelaria, who claims that this technique enables performance close to the performance specs of the NAND media.

Specifically, the company claims performance of more than 100,000 I/Os per second (IOPS) of sustained, random I/O with 4KB block sizes and a 70/30 read/write split. Other performance specs include a latency of 0.1 milliseconds, and bandwidth of 1.7GBps (internal to the chassis).

The other major problem with MLC flash is wear-out. For example, SLC is rated at about 100,000 cycles per cell, while MLC is rated at only 10,000 cycles/cell before the cell becomes unreliable (and that may go down to 2,000 to 4,000 writes/cell with smaller die sizes).

To address the wear-out issue, WhipTail uses a technique called linearization, which essentially entails writing forward across the disk and not revisiting blocks until the entire array has been utilized. This not only decreases wear on the media, but also increases performance. Working in conjunction with linearization, a defrag process ensures that there is always a minimum amount of free space. This technique also works in conjunction with the drive’s wear-leveling algorithms.

The company’s internal tests indicate that if you rewrite an entire array once a day, the device will last seven years (or longer than most other components in the storage hierarchy).

You can get the details on these two techniques on WhipTail’s web site, as well as details on their products, but what about pricing?

Candelaria contends that WhipTail “provides tier-0 [SSD] performance at the price of tier-1 arrays.”

Well, a 1.5TB WhipTail SSD array is priced at $46,000 retail; a 3TB version at $75,600; and the new 6TB configuration, introduced this month, at $122,500.

Summit NJ-based WhipTail was spun out of TheAdmins, a reseller, early this year and has been working on its SSD technology since late 2007. Its first product went GA in February. The company sells through resellers, with eight VARs signed up so far.

For more information on SSDs, see InfoStor’s SSD Topic Center.

And if you’re really interested in solid-state technology, consider attending the Flash Memory Summit, August 11—13 at the Santa Clara Convention Center.

Intel slashes SSD prices

July 23, 2009 –There are still some issues that need to be ironed out with solid-state disk (SSD) drives (e.g., reliability and endurance), but the biggest problem -- and gating factor to adoption -- has been the outrageous price of these devices.

One way to reduce prices is to use the less expensive multi-level cell (MLC) NAND flash technology, as opposed to the more expensive, reliable and durable single-level cell (SLC) technology. But at least for enterprise-class applications, that requires improvements in either controller and/or software technology (which I’ll blog about in an upcoming post).

Another way to reduce SSD prices is to go with a different manufacturing process. That’s what Intel announced this week for its X25-M (Mainstream) line of SSDs, which are admittedly designed primarily for desktops and laptops as opposed to enterprise arrays and applications.

Intel claims a 60% price reduction due to moving from a 50-nanometer manufacturing process to a 34nm process (smaller die size), and a quick price check seems to legitimize those claims.

For example, the 80GB X25-M SSD is channel-priced at $225 in 1,000-unit quantities, a 62% reduction from the original price of $595 a year ago. And the 160GB version is priced at $440, down from $945 when it was first introduced. Both of those SSDs come in a 2.5-inch form factor, with a 1.8-inch version, the X18-M, due in August or September.

Intel claims performance of “the same or better” compared to the 50nm predecessors, citing up to 6,600 I/Os per second (IOPS) on 4KB write operations, and up to 35,000 IOPS on read operations. The company also claims a 25% reduction in latency, to 65 microseconds.

Calculated on a cost-per-GB basis, SSDs are still way more expensive than traditional spinning disk drives, but SSD price wars should come as good news for users with the need for speed.

For more info on Intel’s SSDs, click here.

For general information and news, visit InfoStor’s SSD Topic Center.

We need standards for SSDs

The Storage Networking Industry Association (SNIA) recently announced that it has formed a Solid State Storage Initiative (see SNIA launches SSD initiative). In addition to the SNIA's normal activities such as evangelizing, proselytizing and cheerleading, the SSSI will contribute to standards relating to solid-state disk (SSD) drives.

What we really need here from the SNIA are standards that users, integrators and OEMs can use to compare SSDs, as well as SSDs vs. traditional hard disk drives (HDDs). At a minimum, this standard, or standards, would address performance, providing apples-to-apples metrics to enable comparisons of vendors' performance claims.

But in the case of SSDs, the metrics would have to go way beyond that. For one, they would have to include capacity and price. This would approximate what we get from the Storage Performance Council's SPC benchmarks.

However, the SSD metrics should also encompass durability/reliability and even power consumption. I doubt that it would be possible to come up with a single metric that measured IOPS/$/GB/watts, but the industry will desperately need at least a series of metrics to enable users/integrators/OEMs to make sense out of the nonsense that currently dominates in marketing materials.

The hottest technologies at SNW

In my last blog I promised to reveal The Hottest Technology at last week's Storage Networking World (SNW) show but first, the runners up.

5. Solid-state disk (SSD) drives. There was a lot of talk about SSDs at SNW (mostly in the context of the raging SSD-vs.-HDD debate), but there weren't many product introductions at the show, although Intel did announce production shipments of its new line of enterprise-class flash drives, which will eventually spur further price erosion in this yet-to-get-hot market.

4. Cloud-based storage. Again, a lot of talk, but few vendors. The only cloud storage vendor I met with was Nirvanix, although this services category is expected to grow rapidly over the next couple quarters. For more on this subject, see "What is cloud-based storage?"

3. Storage efficiency technologies, most notably data de-duplication and thin provisioning. Data de-duplication for secondary storage is becoming widespread, but there was a lot of talk at the show about data de-dupe for primary storage from vendors such as NetApp, Storwize, Ocarina Networks, and there will be much more to come.

2. Server virtualization. This is clearly the dominant IT trend, but there are so many storage technologies focused on optimizing virtual servers that it didn't quite make the top of my list.

And the winner is . . .

Fibre Channel over Ethernet (FCoE). This is admittedly an odd choice because, in terms of end-user adoption, FCoE may be years away, but

For one, FCoE was the topic of the only major press conference at the show (hosted by QLogic, Cisco, NetApp and VMware).

Two, a number of vendors made FCoE product -- or at least product certification -- announcements at the show (albeit prototypes in most cases), including EMC, NetApp, Brocade, Emulex and QLogic.

And three, FCoE will provide controversy (as in 10Gbps iSCSI over Ethernet vs. Fibre Channel over Ethernet) for years to come.

I'll address the controversial side of FCoE in my next blog.