[Planetlab-devel] Re: DRL test sites

[Marc Fiuczynski]

Hi Ken,

> On Jan 21, 2009, at 5:08 PM, Kenneth Yocum wrote:
>> If you (or Stephen) can give us the tests and configurations (I'm  
>> assuming they're straightforward), we would be happy to reproduce  
>> and debug in our local cluster.   Off hand, modern boxes should be  
>> able to do 900Mb/s without seg offloading, especially if all  
>> they're doing is tx'ing.

The basic test we run is a basic TCP xput using something like iperf,  
netperf, etc., it doesn't really matter.  The hardware configuration  
is basically two machines connected via a GigE switch.  You should be  
able to reproduce the problem by running our latest kernel (which you  
may already have), install it on two machines, apply the conventional  
htb settings we use.  In case you don't have our latest kernel, you  
can grab the i686 from:
http://build.planet-lab.org/planetlab/f8/planetlab-f8-linux32-4.2- 
branch-2009-01-22/RPMS/i686/kernel-2.6.22.19- 
vs2.3.0.34.32.planetlab.i686.rpm
or for x86_64:
http://build.planet-lab.org/planetlab/f8/planetlab-f8-linux64-4.2- 
branch-2009-01-22/RPMS/x86_64/kernel-2.6.22.19- 
vs2.3.0.34.32.planetlab.x86_64.rpm

Just install this kernel on one node.  What you'll find is that it  
can receive without problems. However, if its the sender it gets  
throttled.  Only after we nuke the htb with something like:

tc qdisc del dev eth0 root

then it seems to work again.  Note that with an older kernel that  
does not have your patch we do not observe this problem.

Hope this is sufficient.  If not, then Stephen Soltesz will likely be  
able to give you more details of the configuration.

Cheers.
Marc