Hello,
I know that you have reduced netdev_max_back_log to 8 to reduce the latency. But
I think that you should know that I had to set it back to 32 to avoid ethernet
frame losses on a 3GHz P4 using e1000 (eth0/threaded=1).
My app is still the one with a 20Mbytes/s flow of data and some signal
processing work. That's only 20% of max Gigabit Ethernet load and 40% of CPU
time. Not really hot stuff.
So, this were my 2 cents to help you decide how to balance latency and
throughput. I hope that you can find a good balance or find a way to allow
people to easily choose their own balance (e.g. maybe there should be a list of
parameters impacting latency in the Documentation directory).
please keep on the good work and thanks a lot !
P.O. Gaillard
* P.O. Gaillard <[email protected]> wrote:
> I know that you have reduced netdev_max_back_log to 8 to reduce the
> latency. But I think that you should know that I had to set it back to
> 32 to avoid ethernet frame losses on a 3GHz P4 using e1000
> (eth0/threaded=1).
could you try the -Q6 patch? That leaves netdev_max_backlog at the
default value of 300 and uses another method to break up the networking
codepaths, without any bad tradeoffs.
Ingo