OpenFOAM machine loads for parallel runs
I have observed that when I use multiple CPUs of the same node is the machine load is quite high (higher than number of proc). In comparison, CFX usually does not behave like this. I observed this pattern with different OF solvers: simpleF, turbF, dieselF, etc.
-When running a job with 4 processes on a Quadcore machine, the load it higher than 4 (~5).
-Even on dual cores load is typically higher ~2.7
(all AMD Opteron machines & used exclusively for OF)
When I use a different (master) node to start a run, then slave nodes have "normal loads = num of Proc".
To avoid this overloading issue, I am using one CPU for mpirun (i.e. Running with #CPU-1 --- 3 proc on quadcore)
It seems that mpirun is responsible for this extra load. Does this mean that OF is not using the CPUs in optimum manner? My guess: when load is more than 100% it is not optimum.
A load average slightly higher than one is not necessary very bad. I think a better metric is CPU time vs. walltime. If walltime is not roughly equal to (CPU time / NCPUs), then your parallelization is not very good.
|All times are GMT -4. The time now is 20:33.|