[Spread-users] Tuning spread's performance for multicast
ciprian at jhu.edu
Fri Apr 4 17:00:16 EST 2003
GA> I got it to work and tested it on two machines on the same
GA> subnet and the performance was underwhelming.
GA> I achieved sustained throughput of 40KB/s for data sent using
GA> multicast with RELIABLE_MESS and FIFO_MESS protocols.
GA> The machines both have 100 Mb cards.
GA> FTP achieved 600 KB/s.
Without having looked at your test code I am first surprised by your
low throughput that you receive using just FTP. When I ftp a file on
our local 100Mb network I reach transfer speeds varying between
5000KB/s and 8500KB/s. On 10Mb networks you should still get around
1000KB/s with ftp. So I think there is either a typo in your numbers
or some major problem with your network.
Ignoring that, the multicast problem may be caused by the fact that
multicast doesn't work properly in your network. If that is the case,
in a two-daemon scenario like the one you use, RELIABLE multicast messages that
do not reach the destination are retransmitted as unicast. The unicast
will succeed, but the performance is severely affected because each
message is sent at least twice.
The quickest way to test this is to run spread on 3 machines using
multicast. If indeed multicast is not working in your network, you
should notice that no message will actually go through (since in the 3
machine scenario the retransmissions will also use multicast and
therefore fail again).
More information about the Spread-users