On Mon, 2010-02-22 at 07:11 +0100, sascha wrote: > The access time for 4 usb sticks is already 0.3ms. Their combined size is > 64GB. Take 32 of these groups and you get down to 0.01ms (if it scales > indeed). > What i was able to test was that it scales up to 16 devices. I got 0.08ms > access time with this configuration IIRC. > Even if you need 2 or even 4 distinct USB host adaptors to connect 128 > devices, you still end up faster and cheaper than with SSD. > Besides that, those PCIe SSDs are still more expensive than 2.5inch SSDs. > And we certainly do not need hundreds of megabytes per second bandwidth. > The write performance is also of no concern. > > As for the optimization of the USB bus, that comes for free. If you start > 4 CPU threads and manage to make each thread access the data blocks of > a single device, then the access time improvements scale linearly with the > number of devices.
Nice! It seems that USB mass storage should have us well covered for 5K requests a seconds :-) I did some further experiments on a single stick, and got access times down to 0.4ms when accessing the block device directly. (Which means we do away with the filesystem altogether.) I added the test code to svn: tmto-svn/tinkering/various/speedy.cpp typically you have to do "sudo time ./speedy /dev/sdX" and get timing figures for 10000 random reads. (Just under 4 seconds in my test) F _______________________________________________ A51 mailing list [email protected] http://lists.lists.reflextor.com/cgi-bin/mailman/listinfo/a51
