On Thu, 7 Aug 2003, Yedidyah Bar-David wrote:
> ... On the last runs it never went below 42MB/s.
I seem to be having something hard limiting throughput at 20MB/sec.
When you look at the HD-Tach graph, it's plain flat. The drive or
striped-drives want to go higher, something, presumeably the controller,
won't let them. If your setup goes above 20MB/S as you have indeed
described, you definitely don't have my problem.
> I can try to make a stripe/mirror/other stuff on it - it's not yet
> production.
I don't think that playing with your setup would shed light on my problem.
However, if you can, try sticking the 4H in a PC,
hook up a single harddrive or two to it and see if it bumps into some bus
limit like 20MB/S or 40MB/S. This is especially easy to see on HD-Tach,
because a normal graph of bandwidth (y-axis) vs. cylinder (x-axis) is not flat,
whereas a bus-bottlenecked harddrive is dead flat.
> The short answer: I guess 45MB/s is the maximum of each single disk,
> and raid5 did not give better results on the maximum. Are you sure
> a stripe of two should give twice the speed of one? Do the specs
> say so?
The short answer: Yes, and No.
The long one:
Striping in _software_, both in Windows 2K/XP and on Linux, gave me
roughly 90%-100% of the combined bandwidth of the drives involved, given that
neither bus (the SCSI or the PCI) or hardware imposed any limitation. For
example, two Good'ol Adaptec 2940UW's, which should technically go 40M/S,
impose a hard 33M/S limit. The drives are faster than that.
Two cards with a harddrive hooked up to each stacked nicely into a 67M/S,
reported by HD-Tach, Sandra, and real-live-testing-with-really-big-files.
IBM Specs just mention that if I stick it in 33/32, I'll have a PCI bus
limit. But that's nowhere near 20M/S.
btw, Winblowz has the disadvantage of not being installable on its own
software RAID. Lamers. Yes, I want my SWAP and /var equivalent to live on
my RAID.
The most likely _assumption_ is that the controller, while able to work on
33MHz/32bit PCI, is not only limited by the PCI bus bottleneck, but also
won't go past one of the 20MB/S SCSI standards (Either SCSI-3 WIDE over 16bit or
UltraSCSI Narrow over 8 bit), so it degrades itself to a
very primitive SCSI standard. This is the smell of things as I'm picking
it up. Hope I'm wrong... Ideas anyone?
--
Miki Shapiro <aris at pharoe dot com>
Unixophilic Software Developer
---------------------------------------------
Tel: +972-(56)-322433 ICQ: 3EE853
---------------------------------------------
"There are two major products that come from Berkeley:
LSD and UNIX. We don't believe this to be a coincidence."
-- Jeremy S. Anderson
=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]
