cvs commit: src/sys/sys mbuf.h src/sys/net if_ethersubr.c src/sys/dev/mxge mxge_lro.c

[Andrew Gallatin]

Andre Oppermann writes:
 > Andrew Gallatin wrote:
 > 
 > > FWIW, LRO triples receive performance for standard frames (3.xGb/s ->
 > > 9.3Gb/s) on decent hardware.
 > 
 > Nice to see that.  The problem with LRO at the moment is that it only
 > works on short RTT links (<1ms) because the TCP stack doesn't do ABC
 > yet and growing the send window with a LRO receiver is going to be
 > painfully slow as the RTT goes up.

FWIW, it is actually seems better than that on a pair of dual-package,
quad core 2.66GHz Xeon X5355s:

dust02% netperf243 -P0 -C -c -Hdust01-m
 65536  32768  32768    10.00      2373.44   7.77     21.41    2.146   5.912  

< enable lro on dust01 >
dust02% netperf243 -P0 -C -c -Hdust01-m
 65536  32768  32768    10.00      9271.76   9.84     22.22    0.695   1.571  

 > Lets add the interface capabilities flag for LRO including the ifconfig
 > support and be done with this episode.

OK.  I'm about out of time for today.  If you haven't beaten
me to it, I will try to do it tomorrow or the next day.
I assume you're willing to review a patch?

Drew