Debating the merits of SSDs
- 14 July, 2008 10:15
Would you pay several times more for a technology that yields only dubious performance advantage? How about if that technology is experiencing a high rate of product returns from early adopters?
That's the central question regarding flash SSDs (solid state drives) as Joel Hagberg, vice president of business development at Fujitsu, sees it.
According to Hagberg, the much-hyped advantages of SSD over CSDs (conventional spinning drives) -- better performance, improved reliability, and lower power consumption -- are minimal and certainly do not justify adding $1,000 to the price of a notebook.
In fact, as Hagberg explains, Fujitsu customers have returned mostly negative feedback on SSD laptops, having found no significant performance advantage nor much improvement in battery life.
"The LCD panel, the CPU, and the DRAM are the three power hogs in a notebook, and those are still there with a SSD," Hagberg says. "When we did the measurements in our labs we saw a 10- to 15-minute improvement with the SSDs, not enough to warrant the price increase if you are not getting the performance increase."
Hagberg does, however, concede that SSDs provide improved reliability when compared with CSDs. And though he is not convinced of the proclaimed performance advantage of current SSDs, he also admits that there are areas -- random reads, for example -- where they are faster than CSDs.
Perhaps this means Hagberg might see a better fit for SSDs in a corporate setting, where customers are less sensitive to price differences if there is a strategic advantage to gain.
"Storage arrays, SAN or NAS, do more than random reads. For example, you have RAID 5 striping happening, where what you need to do to have parity is read-modify writes," Hagberg says.
As a reminder of what Hagberg is referring to, check out Figure 3 of sr5tech.com this diagram. When a block of data in RAID 5 is updated, the storage subsystem needs to read both data and parity blocks, update them, and then write them back. In essence, two read and two write operations are required to update a single block of data.
"Then if you have to do a wear-leveling algorithm for writing, you introduce a level of complexity that reduces those I/O numbers significantly," Hagberg says.
That wear-leveling argument is the ultimate weapon that SSD detractors such as Hagberg throw at their opponents. In essence, the memory cells that flash-memory chips and SSDs are made of can sustain only a finite number of erase-write cycles. There seems to be a consensus around that number: It varies between 100,000 and 1 million cycles, depending on the "quality" of the flash memory used.
Regardless of which number we choose, to extend flash memory's ability to work without errors, manufacturers must implement a wear-leveling strategy -- in essence, distributing writes across cells so that no cell will be forced to an early failure.
If you are interested in learning more about wear-leveling algorithms, this presentation in PDF format from a university in Taiwan is useful, and in English.
But the question remains, is Hagberg right about SSDs? In a word, no, although he does raise some good points. However, I don't think that the clash between wear-leveling and RAID creates an insurmountable problem. For example, this vendor has been offering a RAID card that mounts flash SSD (PDF file) for quite some time.
Nevertheless, at the current state, many customers will probably agree that SSD can be a little pricey for their taste. But there is a range of customers that will appreciate the additional performance and will gladly pay a premium to get it. After all, many people travel first class, and they love it.