Subject: Re: de0/pciide0 performance conflict?
To: None <port-i386@netbsd.org>
From: Allen Briggs <briggs@ninthwonder.com>
List: port-i386
Date: 10/17/1999 19:40:29
I have looked a little bit more into this.  For those who don't
recall...  I have two systems ('canolog' and 'backup').  Both are
AMD K6/233 systems with Apollo VIA3 chipsets, 128MB of RAM and de0
(21140A) attached to a Bay 350T switch.  The two systems are
configured at 100baseTX full-duplex and the switch and ifconfig
both report 100baseTX full-duplex.  'canolog's disks are wide SCSI
on an Adaptec 2940UW.  'backup's main disk is a UDMA/33 EIDE (PIO
4, DMA 2, UDMA 2).

When I ftp a reasonably large file (base.tgz) from canolog to backup
with the default configuration, I get pathetic ( < 100 K/s ) performance.
If I ftp the same file to /dev/null, I get much better (3+ MB/sec)
performance.  This suggests to me that the problem has something to
with the disk on backup or the interaction of the disk with the network
since the disk seems to be OK the rest of the time.

I have also run a few other tests with different kernel configurations.
I tried disabling DMA modes one at a time:

	wd0 on wdc0:			OK -- 3+ MB/sec FTP
	wd0 on pciide0 (UDMA mode 2):	Slow
	wd0 on pciide0 (UDMA mode 1):	Slow
	wd0 on pciide0 (DMA mode 2):	Slow
	wd0 on pciide0 (DMA mode 1):	Slow
	wd0 on pciide0 (PIO mode 4):	OK -- 3+ MB/sec FTP

Does this suggest anything to anyone?
Is anyone running without these kinds of problems with 100Mbit ethernet
and DMA/UDMA disks and the VIA3 chipset?

Thanks,
-allen