Brian's Brain

This blog post references my feature article 'Solid-State Drives Challenge Hard Disks' in EDN's November 13, 2008 edition. It's one of a series of web addendums to the print writeup.

As I mentioned last Thursday in the wrap-up of my initial three-part web addendum to the hands-on feature article, I'd received several 32 GByte SLC (single-level-cell) flash memory-based SSDs from Intel the prior day. I got the chance to run them through my testing suite last night, and below you'll find the results. Click on this link to download the updated ZIP of Sandra report files, and hit this link for an updated spreadsheet consolidation of the data.

By means of introduction, please recall the following portion of my article:

...you need to first understand some high-level differences between SLC (single-level cell)- and MLC-flash technology. [MLC techniques store] 2 bits of information within each flash-memory-array transistor by equating a unique 2-bit data combination with each of four possible charge amounts on the transistor’s floating-gate structure. This arrangement translates to three threshold levels that act as the boundary conditions. Sensing the exact charge value during read operations takes more time than does the traditional SLC approach, which differentiates between only two charge-quantity states—that is, a single threshold level. However, MLC extracts 2 bits’ worth of information at a time—versus 1 bit per transistor access for SLC—counterbalancing the extended MLC-read-access delay.

MLC performance suffers even more during write operations, however. Depositing a precise number of charge electrons on the floating gate is critical to ensuring accurate subsequent read-back results, particularly when you consider that the ensuing data alteration of adjoining transistors can result in minor charge disturbances of the transistor in question. SLC-write operations, conversely, can be more brute force in nature and consequently faster.

In comparing the benchmark stats on Intel's SLC SSDs with those of the company's MLC SSD siblings, I think you'll see that the above forecasts generally hold true. Also be sure to compare the data against that earlier obtained for Samsung and Sandisk's SLC SSDs. it'll be clear, I think, that industry-wide SSD system design expertise remains on a steep learning curve, with newer drives operating faster (albeit with a potential higher power consumption tradeoff) by virtue of factors such as more numerous simultaneous-access channels to the flash memory array coupled with a more intricate flash memory controller, even if the underlying flash memory technology remains essentially unchanged. On that note, I'll clarify that although SiSoftware recently uprev'd Sandra Lite 2009 to the SP1 version, I tested Intel's SLC SSDs using the same initial Sandra Lite 2009 iteration I used with the prior drives.

File Systems Test

Note that the read performance results between Intel's SLC and MLC SSDs are nearly identical. Conversely, the SLC SSDs delivered notably higher write performance...the highest of any drive (HDD or SSD) tested in this study, in fact. Also note (as before with the MLC SSDs) the performance limitation incurred when running the two-drive RAID 0 stripe configuration 'behind' the Silicon Image SteelVine RAID processor in the Addonics enclosure...a bottleneck that was partially relieved by reconfiguring the drives in JBOD mode and instead relying on the RAID 0 software built into Windows Vista, and which was further alleviated by employing the dual ESATA ports built into the Addonics ExpressCard adapter.

Physical Disks Tests

I realized last night that although it's not possible to run Sandra's Physical Disks Write tests on a formatted drive (or multi-drive array), since the drive(s) must be completely empty beforehand, the Physical Disks Read tests complete without problem. I therefore include Physical Disks Read test results for all four Intel SLC SSD Raid 0 configurations. I was not able to go back and comparatively run tests on the Intel MLC SSDs, since these drives are currently being used in the project which will result in my January 8, 2009 hands-on benchmarking cover story on 'value' x86 CPUs.

Removable Storage Tests

My earlier comments on MLC vs SLC SSD performance equally apply to these particular data sets. In general, as I mentioned in the feature article:

As you peruse the results, you’ll likely notice that the drives’ designers optimized firmware and other factors for certain file sizes—to the detriment of others.

To that point, three particular results baffle me. I'm not sure why the read performance of the RAID 0-striped SLC SSDs was poor (both absolutely and relative to other data sets and configurations) when:

• System-connected in the mode where the SteelVine processor configures them as JBOD, with both SSDs within the same enclosure, and with Windows Vista handling the data striping between them, and
• Specifically with the 32 KByte, 256 KByte and 2 MByte data sets.

As these SSDs contain the initial production firmware, I suspect I've found a 'corner case' (combination of access operation and pattern, configuration and data set) where they stumble. I'll pass along my results to Intel, and I'll report back here at Brian's Brain any meaningful feedback I receive from the company.