Analysis

posted on 18 August 2008 17:09

HDD (hard disk drives) reads and writes take about the same time as each other while PC SSDs (the SSDs - solid state drives - without write acceleration) take 100 times as long to write as to read. For the most part an SSD doing only reads rips right past an HDD, while something like an 80/20 read/write balance brings them to comparable speeds. At any ratio beyond that 80/20 mark the SSD underperforms the HDD. This, of course, depends on myriad other factors, but it's a rule of thumb.

 ^{[An analysis contributed by Jim Handy (pictured left) of Objective Analysis.]}

Windows is based on 2 things that pose problems:

1.) It assumes reads and writes take the same amount of time.
2.) It uses 4KB (yes: K) pages in its virtual memory. This is legacy from the first IBM virtual memory systems of the early 1970s (or maybe the late 1960s) when 4KB was a big share of the mainframe's total core memory. This means that every page eviction becomes a discrete 4KB write to disk.

Enterprise-class SSDs have DRAM write buffers that accumulate these small random writes and only pass them to the flash as much larger (groups of writes - consolidated) when there's a need for the DRAM space or the power is lost, so there are very infrequent flash writes. The buffer increases the manufacturing cost of the drive. PC SSDs have no such buffer (except a few, like Micron's new P200) so they choke.

Tuning Windows for a PC-class SSD would involve forcing the O/S to evict larger pages from DRAM and for the other software to postpone disk writes until a larger write page was accumulated. The erase block for NAND chips is 256KB, and this would be the optimum size to get the fastest NAND response. (That's still awfully tiny! Seems like the O/S could cope.)

One of the presenters at last week's Flash Memory Summit (Rob Peglar of Xiotech) traced the disk activity on his PC during boot-up and he said there were 1.8 million I/Os!!! I didn't ask what the read/write balance was, but with that many there must have been a substantial number of writes. Windows is disk-bound, and yes, it could be optimized (probably with great difficulty) for an SSD.

Methinks that the problem of optimizing Windows for an SSD would require a company-wide effort since there has been no single focus on this angle before, so the problem is scattered widely across all the code, shotgun-style. 

It's interesting that Intel's new Turbo Memory is not harping about improvements in boot time. So far that's been about the only thing that SSD makers COULD boast about!

^{[Contributed by Jim Handy, Objective Analysis.]}