One good news I have to share with people:
I had originally mentioned that there were two
TCP-related regressions that need to be addressed:
1. When tested with actual network interfaces
at low numbers of connections.
2. When tested with TCP fusion.
I just tested the TCP fusion case with the stock
kernel on NV 94. It exhibits the same behavior
with similar lockstat, etc. So, there is no
regression to fix. I must have assumed when
I saw the throughput decrease for higher numbers
of connections that my kernel has introduced
it. But even for the stock kernel, the exact same
behavior is seen.
So, we have only one outstanding perf problem.
Issue 1 stated above.
Stay tuned.
Madhavan
