On Mon, Oct 31, 2011 at 8:10 PM, Rick Jones [off-list ref] wrote:

Whether tracked as bytes or segments, my take is that to ask applications to
have to think about another non-portable socket option is ungood.  I would
suggest taking the time to work-out the automagic heuristic to drop the
deferred ACK count on connections where it being large is un-desirable and
then not need to worry about the limits being global.

Your suggestion deserves further investigation, it looks tricky to
find a good heuristic for increasing/decreasing the ACK deferred count.

Given the stack's existing propensity to try to decide when to increase the
window I might even go so far as to suggest the sense of the heuristic be
flipped and it seek to decide when it is ok to increase the number of
segments/bytes per ACK.  To what extent one needs to go beyond what happens
already with the stretching of ACKs via GRO/LRO or if that mechanism can
serve as part of the logic of the heuristic is probably a fertile area for
discussion.

If I recall correctly, in one of your earlier posts you mentioned something
about a 20% performance boost.  What were the specific conditions of that
testing?  Was it over a setup where the receiver already had LRO/GRO or was
it over a more plain receiver NIC without that functionality?

If I remember correctly on the receiver side there was no LRO/GRO, but we
tweaked some of /proc/sys/net/ipv4 parameters (e.g tcp_rmem).
Also, the traffic was highly unidirectional with many clients feeding multimedia
content to a server.

Anyhow, we used our custom kernel which is an older kernel version.
Are there any recommended benchmarks/tools for testing this kind of parameters?

Daniel.