time measure

Brooks Moses

2006-06-25 19:16:39 UTC

Post by wojek
I'm writing a survey of distributed computations in which I want to compare
various algorithms on various architectures. the problem is that I don't
know how to measure their efficency. I can of course use Linux's time, but
the problem is that I get real/user/sys times. Real time would be pretty god
if I have computer/cluster exclusively. User/sys tells a lot, but it doesn't
tell me how long my process was waiting for communication which is a huge
problem on clusters.

I don't think a complete solution is (reasonably) possible.

How would you expect node 2 to be able to tell the difference between
the time it takes waiting for communication from node 1 while node 1 is
busy with your calculation, and the time it takes waiting for
communication from node 1 while node 1 is busy with someone else's
calculation? Or, for that matter, how would you count the time when
node 2 would have been waiting for communication from node 1, except for
the fact that it's been delayed by doing someone else's calculation instead?

If you're using synchronous communication, you might be able to put some
timing loops within the code to measure the time spent waiting on those
particular commands. (And, for all I know, your communication library
might have such included already -- in any case, the method for doing it
is likely to depend on exactly what communication library you're using.)
My guess is that that's about the best you could do easily, and it's
not likely to be independent of cluster load.

- Brooks

--
The "bmoses-nospam" address is valid; no unmunging needed.