On Tuesday, 2 December 2003 at 10:26:08 +0100, Sander Smeenk wrote: > Quoting Greg 'groggy' Lehey ([EMAIL PROTECTED]): >>>> plex org striped 3841k >> >> You should choose a stripe size which is a multiple of the block size >> (presumably 16 kB). Not doing so will have a minor performance >> impact, but nothing like what you describe here. > > I might have misunderstood, but on the vinumvm.org website there is > quite a comprehensive discussion on stripesizes and they conclude > that larger stripesizes help increase throughput. They also discuss > the size not being a power of 2, because that might cause stripes to > end up on the same disk, which decreases performance...
That's all correct, but unfortunately the discussion isn't comprehensive enough. Since transfers tend to be on block boundaries (but there's no requirement), it's a good idea to have all blocks on a single platter. Otherwise transfers can get broken into two, with minor performance implications. I'm currently tending towards a stripe size of 496 kB. >>>> ahd1: PCI error Interrupt >>>>>>>>>>>>>>>>>>>>>> Dump Card State Begins <<<<<<<<<<<<<<<<< >>>> ahd1: Dumping Card State at program address 0x94 Mode 0x22 >> This is possibly related. Does it happen every time? > > It did, until I compiled a new 4.9 kernel from the 4.9-RELEASE src/ tree > from CVS. (Thanks to Scott Long for pointing that out). > The driver for aic7xxx cards was fixed, and now the message is gone, and > the system is once again stable. Ah, OK. >> The first thing to do is to find whether it's Vinum or the SCSI disks. >> Can you test with a single SCSI disk (of the same kind, preferably one >> of the array) instead of a single IDE disk? > > I did some tests, this time with Bonnie++, on vinum, scsi, ide, and > vinum with big stripes and small stripes. I'm busy comparing them ;) > > But still, I doubt if bonnie++ is a good test, and I have a hard time > interpreting the results. I can publish the results somewhere, in a > while. Yes, bonnie++ tests the entire system, not just the disk. Try benchmarks/rawio. Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers.
Description: PGP signature