7.0-RC1 bge network performance issue?

[Gregory Wright]

Hi JD,


On Dec 30, 2007, at 9:31 AM, JD Bronson wrote:

> I am not sure how or where to ask this so I will start here..
>
> I have an IBM P4-3.06ghz machine with dual 10/100 BGE nics and 1GB ram
> with ATA133 IDE drives.
>
> I have noticed the following and wondering how I can troubleshoot  
> this or what's changed.
>
> Using ANY 6.x branch, including 6.3-RC1 I have no issues at all.
> Using the same hardware/machine exactly and using 7.0-RC1
> I am seeing terrible network writes into the machine.
>
> my bge nics are set to 100FDX and incoming FTP rates are dismal
> at around 2-3MB/sec max.
>
> FTP outbound rates OUT of the box are at full 11-12MB/sec.
>
> Using 6.3-RC1 - traffic is the same in or out of the box.
> (I am not using pf)
>

One thing to check is whether you are getting duplicate ACKs out of  
your box.
A good way to check is to set up netperf client on a spare machine and  
a netperf
server on your P4/bge box.  Do a TCP test while looking at the packets  
with
Wireshark (or your favorite packet logger).

The duplicate ACK problem seems to be caused by a firmware bug on some  
models
of the bge chipset.  It's not just a FreeBSD problem, it was seen on  
Gentoo linux
as well.  (Apparently later versions of the gt3 driver --- the linux  
equivalent of bge ---
fixed this problem, but it is not immediately apparent what the magic  
change was.)
I saw the duplicate ACK problem on both 6.x and 7.0-BETA2, but it  
seemed a good
deal worse with 7.0.  One interesting thing on my dual Opteron box was  
that the
problem didn't happen when running at 1Gb/s, it showed up when I ran  
at 100 Mb/s.

I suspect some subtlety in initializing the chip is behind the  
duplicate ACKs.  Some
people have suggested TCP segmentation offloading (TSO) could be the  
root cause.  You could
also  check if TSO is enabled, and if it is, disable it with ifconfig  
(8).

Best Wishes,
Greg

Gregory Wright
Antiope Associates LLC
18 Clay Street
Fair Haven, New Jersey 07704
USA

gwright at antiope.com

1 (732) 924-4549
1 (732) 345-8378 [fax]

> If it's any help, I have noticed the exact same issue with OpenBSD  
> since anything newer than 3.7 and my posts to that group with this  
> issue have basically been unresponded to.
>
> I would love to keep using FreeBSD but cant with this issue.
>
> NetBSD 4.0 does not exhibit this behavior at this time.
>
> Any thoughts? Major changes on bge? Anything?
>
> It simply cant be my hardware as this is not an issue under the 6.x  
> branch.
>
> Thanks,
>
> -JD
>
> _______________________________________________
> freebsd-current at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe at freebsd.org 
> "
>