[Bioclusters] gigabit ethernet performance

Joe Landman bioclusters@bioinformatics.org
Fri, 09 Jul 2004 22:33:35 -0400


I have also seen in the 60 - 80 MB/s for real applications on e325's
using the bcm5700 drivers (the tg3's don't work so well).

Generally there are many reasons why gigabit performance can be bad. 
Switch performance is one of them.  Network settings are another.

The original query was on NFS, and how the Xserves were getting 12ish
MB/s on GB.  This sounds suspiciously like someone somewhere is locked
into 100 Mb/s mode on a port they think is 1000 Mb/s.  When running at
full tilt, a good NFS server implementation on 100 Mb links can source
about 11.7 - 12 MB/s.  You would see similar performance from rcp in
this case.

If this is not the problem (and I recommend sanity checks, as in
checking what both sides of the connection report, as many switches are
known to autonegotiate incorrectly), start looking at things like MTU
(can you use jumbo packets?), TCP based NFS, larger read/write sizes,
...

Some of the IDE RAID systems we have set up have been able to
sink/source upwards of 60 MB/s without working hard at tuning, and we
have seen a sustained 70+/- MB/s for over a 2 day run at a customer
site.  We cannot speak to the Xserve as we don't normally use or spec
it.

Look in the usual spots, and make sure that you leave nothing to an
assumption.  It is possible that you will run into driver issues,
network stack implementation, bad switches...  

Joe

On Fri, 2004-07-09 at 21:46, Farul M. Ghazali wrote:
> On Fri, 9 Jul 2004, Tim Cutts wrote:
> 
> > ~30 MB/sec sounds about right to me.  We get 36 MB/sec between our
> > AlphaServer ES45 boxes; that's GBit ethernet, and HP StorageWorks HSV
> > RAID controllers; there's no way the disk is the limiting factor in our
> > set up - we can get about 200 MB/sec on the HSV controllers.  I think
> > the low 30's is about what GBit can sustain.
> 
> 30MB/sec is pretty bad for gigabit if that's the case. I've used netperf
> and gotten 70-80MB/sec on my IBM x325s, but I don't have fast enough disks
> to really test the system.
> 
> Does the switch come into play in this? I've used el-cheapo Netgear
> switches and we are looking into geting some demo Foundry swtches to test
> out.
> 
> _______________________________________________
> Bioclusters maillist  -  Bioclusters@bioinformatics.org
> https://bioinformatics.org/mailman/listinfo/bioclusters
-- 
Joseph Landman, Ph.D
Scalable Informatics LLC,
email: landman@scalableinformatics.com
web  : http://scalableinformatics.com
phone: +1 734 612 4615