Subject: Re: de0 on 164LX gives poor throughput
To: Bernd Ernesti <netbsd@arresum.inka.de>
From: Thor Lancelot Simon <tls@rek.tjls.com>
List: port-alpha
Date: 04/09/1999 13:49:52
On Fri, Apr 09, 1999 at 06:40:39PM +0200, Bernd Ernesti wrote:
> On Fri Apr  9 18:05:00 1999, Paul H. Anderson wrote:
> > 
> > 
> > Today is yet another day in alpha land when 1+1 doesn't equal 2.
> > 
> > I've got several racks full of 164LX machines running netbsd 1.3I from
> > 19990122 snapshot.
> > 
> > Each machine HW is: alpha 164LX Dec motherboard, 533MHz 21164, de500
> > 10/100, ide boot disk, adaptec 2940UW, 1GB RAM, two 10K rpm drives, serial
> > console.
> > 
> > UDP from an alpha to other alphas seems ok (NFS to a server gives about
> > 3-4 MB/sec for reading files).
> > 
> > TCP streams via FTP seem FUBAR.  In pretty much any combination that we've
> > tried, an alpha talking to an alpha gives about 50-100KBytes/second
> > throughput - about 1/100 of what it should.  Same HW ifconfig'd to 10baseT
> > gives 1MB/sec, i.e. full 10baseT throughput.
> 
> Please show us the ifconfig de0 output.
> 
> Do you have an switch which supports full-duplex, if yes then and try to use
> 'media 100baseTX mediaopt full-duplex' and not autoselect.

DO NOT DO THIS!  Hardwire the _switch_ end of the link to full-duplex, and
let the card autoselect.  I've been testing for about a week now with several
"de" cards and switches and if the card is set to full-duplex, our driver
seems to completely screw up the negotiation with the switch somehow leading
to awful performance.  Leaving both ends on "auto" of course doesn't always
work well, either -- Cisco switches are notorious for screwing this up --
and all that I've found that *does* work reliably in every case is switch
hardwired to "full", card on "auto".

I guess YMMV, but I've tested with enough switches and cards to have a pretty
good idea what the score is...

Our "de" driver should be defenstrated.