Hi,
some time ago I ran the "1 vs p ping-pong" experiment
on our cluster: p processes are sending packets to one
central process, who does nothing but to return each package to
the sender. The whole process stops as soon as the central
process has received a given number of packets, and each sending
process then reports how many packets it has exchanged with the
central process.
This sort of experiment is used to demonstrate the lack of fairness
in MPI message handling. Even in a perfectly homogenous system, the
number of packets that contribute to the global exchange may vary
extremely from process to process. The MPI standard does not
guarantee a fair treatment of all messages that arrive at the central
node. And this is precisely what I saw some time ago in my experiments
with LAM 6.5.1 (on 10 nodes, homogenous system).
When I repeat the same measurement on the same system
but with LAM 7.0.6 the fairness seems to have improved considerably.
Question for the LAM development team: Do you have made changes along
the way from 6.5.1 to 7.0.6 to improve fairness in message handling?
What algorithm do you actually use to prevent pending messages from
starvation?
Cheers,
Stephan
--
Stephan Mertens @ http://www.uni-magdeburg.de/mertens
Supercomputing in Magdeburg @ http://tina.nat.uni-magdeburg.de
|
