Date: Wed, 2 Apr 2003 18:33:50 +0200 From: Maxime Henrion <mux@freebsd.org> To: Fred Clift <fclift@verio.net> Cc: alpha@freebsd.org Subject: Re: call for testers: busdma-ified fxp(4) driver Message-ID: <20030402163350.GI1750@elvis.mu.org> In-Reply-To: <20030402090802.Y82002-400000@vespa.dmz.orem.verio.net> References: <20030401134113.GC1750@elvis.mu.org> <20030402090802.Y82002-400000@vespa.dmz.orem.verio.net>
next in thread | previous in thread | raw e-mail | index | archive | help
Fred Clift wrote: > On Tue, 1 Apr 2003, Maxime Henrion wrote: > > > > My fxp(4) patch should now be nearly commit ready. > ... > > > someone with an alpha box and a fxp(4) card could test this patch > > > > Sorry to follow-up to myself, but this patch had some bogus and > > irrelevant diffs in it. I'm attaching a fixed one now. It can also be > > found at http://mu.org/~mux/patches/if_fxp.patch. > > Well, the patch seems to work just fine - better performance now in fact. > > I did a bunch of scp/ftp type stuff of big and a bunch of small files just > to ensure everyting was working and then I used /usr/ports/net/netperf to > try and characterize performance a bit. After you install the port there > is a script /usr/local/netperf/snapshot_script that tries to give a > reasonable snapshot of network performance under a variety of conditions, > and with a few different parameters... The target machine was a 400Mhz > pentium II box (sadly, this and the alpha are the two fastest machines I > own...) with '4.7-STABLE FreeBSD 4.7-STABLE #5: Mon Nov 11 12:15:39 MST > 2002' (+ a few security patches) with a newer fxp card > > fxp0@pci0:9:0: class=0x020000 card=0x000c8086 chip=0x12298086 rev=0x08 > hdr=0x00 > > Both of these boxes plugged into an SMC 10/100 switch... I would have > just used a cross-over cable except I forgot to bring it with me and > didn't feel like making one. > > Attached to this email you'll find the output of snapshot_script for the > old driver (about two weeks old cvsup), the output of the modified driver, > and the output of the script run on the integrated dc NIC, just for fun > and comparison. > > Some sample results at one particular set of parameters: > > before: throughput in 10^6 bits/sec 39.18 > after: throughput in 10^6 bits/sec 94.11 > (dc): throughput in 10^6 bits/sec 70.97 > > > Honestly I'm suprised there is so much difference between the before and > after and I'm doubting my testing methodology -- perhaps I had the duplex > set wrong on the alpha on the 'before'? shrug. I didn't shut down, or > change at all, the target test box between the two tests. At any rate, the > patch seems to work fine on my alpha on the one card I had time to test > (older, slightly larger formfactor etherexpress pro 100). If you can wait > till tomorrow, I can probably squeeze in time for testing one or two other > fxp hardware revs I have... > > The card I did test on shows up as this: > > pciconf -v -l > ... > fxp0@pci1:8:0: class=0x020000 card=0x00098086 chip=0x12298086 rev=0x04 > hdr=0x00 > vendor = 'Intel Corporation' > device = '82557/8/9 EtherExpress PRO/100(B) Ethernet Adapter' > class = network > subclass = ethernet > ... Great, many thanks to you for testing this! I'm also very surprised about that huge performance improvement, I just can't see how this patch could improve performance :-). If you feel like and have time to do some more performance testing, that would be very interesting. I'm going to commit this patch now. Cheers, Maxime
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030402163350.GI1750>