Appendix L: Solid State Discs Set Off Flash Flood

A Solid State Disc (SSD) mimics a HDD but with non-volatile flash memory replacing rotating platters. Will SSD replace HDD anytime soon? For some applications, it’s happening now; for others, it may never happen. Comparing the two technologies is partially an “apples to oranges” scenario. True, they are both fruit—but the differences are significant even with identical I/O ports (SATA or SAS) and slot form factor.

The key SSD advantages over HDD are as follows: nearly instantaneous R/W access; lower active power consumption by 3–5× and ~40× for inactive state; no moving parts; robust, longer MTBF by a factor of ~3×; no noise; no file fragmentation. Surprisingly, SSD is not far behind HDD in terms of capacity; BiTMICRO Networks offers a 2.5 inch SATA SSD with a capacity of 832GB.

Small data block, read-intensive applications show the most leverage compared to HDD. A report from Burleson Consulting– see www.dba-oracle.com concluded that, for a query-intensive Oracle 9 database server benchmark, SSDbased storage outperformed HDD by the respectable factor of 176. For readintensive A/V applications (VOD server, video server, field camera, etc.) SSD or pure Flash cards are replacing HDD now and will continue to do so as price points improve. For several years, Toshiba has offered the Flash-based ON-AIR MAX video server for broadcast and digital cinema operations.

According to Sun Microsystems, a 146GB disk drive with 15,000 RPM gets about 180 write IOPS and 320 read IOPS, while a 32GB Flash drive gets 5,000 or more write IOPS and at least 30,000 read IOPS. Disk drives ended up costing $2.43 per IOPS (as of late 2008), and solid state drives using Flash cost $0.08 per IOPS. When performance is needed, SSDs will be used. For an example of a hybrid SSD/HDD storage system that supports 288TB HDD and 600 GB SSD, see Sun’s Amber Road Storage 7K series.

However, not all is rosy. SSDs suffer a ~ 100× cost/GB burden in 2008 compared to a SATA HDD and less for a SAS drive. Cost/GB will improve but likely never reach parity. Also, Flash has an upper limit on the number of write cycles. This number is always improving, but 5 million cycles (www.mtron.com) is state of the art in 2009. Most SSD file systems use wear leveling techniques to spread out writes for even distribution. Don’t expect a Flash memory to drop dead after the write cycle spec is exceeded. Rather, write errors will slowly creep in and require strong read error correction methods to hide them. Some vendors use strong error correction to fend off this issue as long as possible.

The read/write operations mix for a SSD can yield drastically different access data rates. For example, a small block, random SSD read may be ~20× faster than for a SAS HDD. On the other hand, for small random writes, SSD is slower by ~15×. Importantly, for a 50/50 small block read/write mix, the HDD performance is ~8× faster! So, the message is clear: small writes can ruin SSD read performance. See Easy Computer Company’s white papers (managedflash. com) for an excellent analysis of this problem and mitigation techniques. So, you should not choose a SSD in place of a HDD without knowing the application scenarios and R/W mix. Something as seemingly innocent as a frequent status log update written to a SSD can ruin the overall performance of a mission-critical application. Designer and user beware.

In the big picture, expect to see select application servers, including some video servers, use a mix of HDD and Flash. Flash offers the best IOPS performance for large block R/W with access speeds exceeding HDD.

L.0 THE MEMRISTOR—A NEW APPROACH TO NON-VOLATILE MEMORY

In 1971 Leon Chua of UC Berkeley (Chua 1971) predicated the Memristor, a nonlinear, passive, two-terminal, circuit element. It is described as the fourth basic type of passive circuit element, in league with the capacitor, resistor, and inductor. It was a hypothetical device for 37 years, with no known physical examples.

In April 2008, a working device with similar characteristics to a Memristor was announced by a team of researchers at HP Labs (www.spectrum.ieee.org/may08/6207). The Memristor may enable a new class of high-density, non-volatile digital memory. When a voltage is applied, the Memristor remembers how much was applied and for how long. Its resistance is a function of the applied voltage. This fact can be used to store one or more bits per element. It could surpass Flash memory as a new, high-density memory element.

References

Chua, Leon O. (September 1971). Memristor—The Missing Circuit Element. IEEE Transactions on Circuit Theory CT, 18 (5), 507–519.