On Thu, 2006年02月16日 at 21:14 -0800, Andrew Warfield wrote:
> > Here's the compulsory benchmark numbers (same machine, uniproc 3GHz
> > Pentium 4), ext2 filesystem in a file in dom0 served to dom1:
> >
> > Current block device:
> > dd if=/dev/zero of=test bs=1M count=1000: 71.5527 seconds
>
>
> By my accounting that looks to be about 14 MB/s off your block device
> using the existing split drivers... If you can't saturate the disk on
> a 3 Ghz box I'd have to think that something has gone desperately
> wrong.
Yes, I think so too. Frankly, I expected my implementation's
performance to be bad, because the hypervisor copies to dom0 userspace,
which then writes to the disk. This is great for implementing COW and
other specials, but not optimal if you're just shunting blocks to/from
disk.
Now, in dom0 this operation takes about 27 seconds (38 MB/sec). My code
gets about 33.7 MB/sec. Now, page-table ops involved in granting
buffers will hurt you some, but I'd expect nice big batches of ops from
this example to amortize that nicely. I cannot explain this level of
pain, but over 10 runs it was consistent.
> I don't doubt that better memory performance can be achieved by
> adjusting event notification frequency/batching -- as you pointed out
> earlier... but block devices are a pretty weird place to try to win. 
> I'd expect the numbers from the two tests you're running to be
> identical -- and bottlenecked on the disk. By a lot.
Hmm, my dom0 only has 128MB of RAM, and my dom1 64MB. Wonder if I'm
getting bad behaviour due to small amounts of memory?
I was just happy to get my code working, but I'll dig deeper next week.
Cheers,
Rusty.
-- 
 ccontrol: http://ozlabs.org/~rusty/ccontrol
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel