Hi,
we are running 2.6.23 (and might go to 2.6.24 soon), but are a bit
puzzled which TCP congestion control to use in this set-up:
Say, we have three switches. Switches A and C are identical with 48 x 1
Gb/s ports and 4 x 10 Gb/s uplinks. Internally, ports 1-12, 13-24, 25-36
and 37-48 are directly mapped to the four uplinks. Switch B in the
center has many 10 Gb/s ports joining these 8 links together.
The RTTs are quite small, ping shows me a maximum of about 100 us.
Scenario 1:
We are using 10 nodes on each port group on switch A and C, thus a total
of 80. When we are now starting netperf with TCP_RR and large MTU sizes
which is the best TCP congestion control mechanism to create a fair
share for each node? In principle there should not be a bottle neck
since it should "just" fit the pipes, however, the switches themselves
insert probe packets to detect possible problems and these tiny packets
might cause havoc from time to time.
Which is the "optimal" congestion control, if there is one. So far we
have done some tests with Reno, BIC, CUBIC and a few others, however,
the results are not really clear and we are lacking experience.
Scenario 2:
Now we are deliberately oversubscribing the switches A and C by putting
11, 12, 11, 10 lines to the four port groups[1]. Also, in this example
we would like to use a congestion control which gives every computer a
share close to what is fair. Anyone with hints, which could be the one?
Thanks for reading this long stuff and more thanks for pointers how to
find an "optimal" solution.
Cheers
Carsten
[1] The reason for the 10 is simply that there is a file server
connected to this port group and we don't want to oversubscribe that
one. Apart from this there are 42 additional computers needed to be
connected as well.
--
To unsubscribe from this list: send the line "unsubscribe linux-net" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
