[Spread-users] Multicast and Performance issues
David Turland
david.turland at shazamteam.com
Fri Mar 15 16:03:32 EST 2002
Hi,
Currently using Spread in earnest and trying to reach the plateau of 1MB/sec
multicast to 36 nodes(Athlon 1.2G,Fast ethernet, switched, local subnet).
(This is the performance I was getting using LAM/MPI and a tree-based
broadcast - If I can't achieve this then I will tolerate the hit of
incorporating failover support using LAM/MPI)
so far topping out at 333kB/sec; whether I choose RELIABLE or UNRELIABLE.
I feel Iam not configuring Mlticast correctly as this value is the same
whether I use a unicast rather than Multicast address in spread.conf
Spread_Segment 192.168.42.255:4803 {
master 192.168.42.91
.
.
.
vs.
Spread_Segment 225.0.1.1:4803 {
master 192.168.42.91
.
.
.
I really know nothing about multicast and feel there is something more I need
to be doing than choosing 'arbitrary' multicast addresses in the spread.conf
incidentally,there is an entry in the /etc/rc.local file pertaining to
multicast:
# Setup the default multicast route
/sbin/route add -net 224.0.0.0 netmask 240.0.0.0 dev eth0
I think I need some more info before I spproach our sys admin guys...
Currently the benchmark I am running uisngs two threads, one for sending and
receiving with synchronisation (cv) between the sending and receiving threads
so that the acks from (one or all ; makes little differnece) of the receivers
are received before the next sending iteration takes place. If I omit this
lock step then the session is closed arbitrarily, I presume because excessive
mesages are stacked up in a buffer.
Interestingly, I played with max packet size, up to 1MB, but with my current,
(non-multicast?) configuration I saw no performance changes.
I love the concept of spread, there are bits of it I am taking for granted
which will take an age in LAM/MPI or RMT or RMTP to implement but I need to
reach the magic 1MB/sec, or beyond
More information about the Spread-users
mailing list