[General boards] [Winter 2021 courses] [Fall 2020 courses] [Summer 2020 courses] [Older or newer terms]

Should we run MPI with -pernode in A0?


#1

After running cdfruptimesort on wolf, I got a total of 22 usable work stations. It is not enough to run 32 threads with -pernode option. However, if we do not run our program in -pernode mode, I think we will not get reliable estimations for the parameters in the problem, as the communication price between different cores in one single node should be lower than that between different nodes. In this case, I think the best option for us might be running our program on wolf in local mode, as wolf have 48 cores available, all communications in our program would be local.


#2

I do not think that -pernode is so critical for the reliability of the results.
At least, I did not see huge variations with it and without it.
But it may depend on the times I chose to run.
My first suggestion is to use the workstations, even if in the core mode,
as the wall clock time will be more reliable on them.
Whatever you do, you need to state it clearly in the assignment.