slacker (s@lack.er) on 8/20/08 wrote:
---------------------------
>Richard Cownie (tich@pobox.com) on 8/20/08 wrote:
>---------------------------
>>I don't see why. USB flash drives are pretty simple and should be quite reliable
>>if they're in a cool machine room rather than in your pocket.
>
>The fact that people 'keep them in their pockets' is the entire reason they're
>perceived as reliable: they go through maybe 100 write cycles before people lose them and buy another one.
>
>>And a system with
>>2000+ disks already has to deal with all the RAID/fault-tolerance gizmology, so
>>that should be able to deal with random failures of individual drives.
>
>You're also talking about cheap MLC flash with an MTBF of less than 10,000 write
>cycles. What's the availability on 2000 chips at 12h/day of full load? What's the
>average read/write bandwidth? What's the cost to the host processor to deal with
>2000 USB controllers? Can you even connect 2000 USB devices to a single system,
>or are you stuck at 127 (not including hubs)?


He's talking about systems designed to run the TPC benchmarks well, not work well in real world use for several years. The configurations being used in TPC benchmarks already bear little relationship to those used in the real world, so why not?

Dunno about USB connection limits, but as the original poster points out, you can get four USB ports on a PCI card for $20, that lets you hook up 500 devices per PCI card.

Anyway, I think the suggestion was somewhat tongue in cheek, poking fun at the silly designs used to run TPC benchmarks that are totally unlike anything anyone would design to run a real-world database (not just in storage configuration, its also true in terms of clustering used, DB layout, etc.)

I think the real problem would come in the per device write latency. Most of those cheap USB sticks allow fewer writes per second than even the slowest laptop hard drive, so thousands of such devices won't get you the IOPS of a couple thousand real disks. There are faster USB sticks out there, but then you are probably no longer at the $3/GB price point.

When Intel announced those SSD drives that sustained 250MB/sec read and 170MB/sec write (using 10 parallel flash channels internally) did they mention any IOPS figures? Any IDF attendees know? That would tell us where flash really is in comparison to the best 15K rpm drives in terms of IOPS. Obviously those drives are only providing 40% of the IOPS that the current TPC leaders need right now, which is why they are overcommitting storage by 2.5x. If Intel's devices can provide 2.5x more IOPS (which would be roughly 1000 IOPS) then if they were able to sell them for 2.5x more per GB than a 73GB 15K rpm it would make sense to switch to SSD for TPC benchmark runs.