Originally posted by: healyp

If I run  several identical instances of CPLEX in single-thread mode I note that the more instances I run the slower the solve time is.  I am curious as to what specifically causes this slowdown.

At a facile level one might say "well, what else did you expect?!" but I'd like to know a little more about what the bottle-neck is. 

Can anybody explain where the contention is, even when the number of instances is smaller than the no. of CPUs available?

Many thanks.

Hi,

do you run all those instances in parallel ?

how do you call them ?

regards

Originally posted by: healyp

Each CPLEX instance is called via the JuMP interface of the julia programming language.  I have a master process that farms out jobs (in this case the very same instance) to a number of pre-specified workers on each remote machine.  I have come to notice that the instance solves fastest when it is the only one solved on that machine.  Assigning 2, 4, ... instances to the same machine causes each instance to be solved slower.  All machines, in a student lab, are identical.

Thanks.

All the instances share the same memory and memory bus. So this is one bottle neck. This also includes the cache and things like cache misses etc.

Also, if you don't use explicit CPU binding, multiple instances may end up on the same CPU, which will of course cause a slowdown as well.