Subject: Re: UVM/other problems for desktop users in current?
To: Peter Seebach <email@example.com>
From: Jukka Marin <firstname.lastname@example.org>
Date: 12/18/2002 09:23:32
On Wed, Dec 18, 2002 at 01:11:42AM -0600, Peter Seebach wrote:
> Anyone seeing this on SCSI? Anyone able to reproduce it on SCSI?
That would be interesting to hear.
> If we're seeing horrible performance botches on one-drive-per-port IDE
> systems (and I am, and you are), but we can't find them on SCSI, we have
> a culprit.
Yep. And I'd appreciate it if someone explained _why_ decent IDE with DMA
is so bad when compared to SCSI. Especially if you only have one device
per port and the raw sustained performance is 30 MB/s or more with only a
little CPU loading.
> Despite all the griping about this, I think we're onto something positive
> that could improve the system substantially.
Absolutely, if we can make the Big Guys admit that there _is_ a problem ;-)
One problem is that the disk operations don't have priorities. If CVS
pushes two zillion disk I/O requests in the queue (I don't know how the
I/O system works, so don't shoot me (yet)) and an interactive process then
requests for one kB, it seems that the two zillion requests are serviced
first and the interactive process just waits.. and waits. And what about
CVS+NFS - neither of them is "interactive", but NFS clearly needs a higher
"priority" to keep the clients happy..