Load balancing CESM2

[Shreya Bhowmick]

Hi everyone,

I have ported CESM 2.2.0 to a Linux server ( Intel(R) Xeon(R) 4310 ) with 2 CPUs(or sockets) each having 12 cores, 2 threads per core i.e. 48 processes. I am running the compset FHIST with irrigation starting from year 1979. Looking at the output files, I find that the 14 model days are simulated in 3 hours (approx. throughput of 0.306 simulated years/day) which is quite slow. I have attached the PE layout. Is there possible way to increase the speed of the run?
Earlier, I used the same architecture for running FAMIPCAM5 compset in cesm 1.2.2 and got the throughput of 0.64 simulated years/day.

Thank you for your time.
With regards,
Shreya.

 

[jedwards]

I'm not sure how this works when you have MAX_MPI_TASKS_PER_NODE=12. If you are running 48 tasks on one node than this value should be 48 as should MAX_TASKS_PER_NODE. In any case I think that you can expect cesm2 to cost about 2x what cesm1 did so your results are not surprising.

 

[Shreya Bhowmick]

Thank you so much!