Subject: Performance weirdness (was: Re: Are there any known problems with uvm or pmap in -current as of about Oct. 16?)
To: Chuck Silvers <chuq@chuq.com>
From: David Johnston <david@canopus.apana.org.au>
List: port-macppc
Date: 11/07/2001 00:25:00
On Sun, Nov 04, 2001 at 11:55:49PM -0800, Chuck Silvers wrote:
> hi,
> 
> matt and I spent a bunch of time working on the NEWPMAP problems today,
> and we found and fixed several bugs.  our stress-test was "make -j 20" in
> a kernel build dir.  when we started this morning, it took about 30 seconds
> to trigger one of the assertions.  a few minutes ago a test run completed
> without a problem after 60 minutes.
> 
> at this point the new pmap appears to be much more stable than the old one,
> so I recommend that everyone turn on "options NEWPMAP" if you haven't
> done so already.
>

I've just built a current kernel with options NEWPMAP turned on, and I'm
seeing a weird performance problem.

My machine is an upgraded 7500. It's got a 300Mhz G3 card and a PCI ATA/100
controller (based around the CMD PCI0649 chip) with an attached ATA/100 disk.

Generally performance with the new pmap is the same as with the old, except
when I transfer a file from an external machine to the ATA/100 disk using
ftp (either using ftp 'put' from the external machine, or by using 'get'
from the macppc machine). For this case I'm seeing terrible performance of
< 500KB/s over 100 baseT, when with the old pmap I would get 5-7 MB/s.

I don't see the performance problem if I ftp to a disk on the machines'
mesh SCSI bus, or if I copy the file to the ATA/100 disk over NFS.

Weird huh? Any ideas?

	david

-- 
David Johnston
Email: david@canopus.apana.org.au	Mobile: 0402 088010