I'm attempting to run a parallel MrBayes analysis using MPI on a Linux machine with 33 processor cores (it's not a cluster, it's just the single multicore machine) but I can't get it to work. It'll start up and run but it just drags. If I run it on a single processor it goes almost twice as fast! Is there some setting I'm missing to let it properly run in parallel?
I've tried both using the bottled version (sudo apt install mrbayes-mpi) and compiling from source with --with-mpi. There are no errors with installation or compilation, but I get the same result either way.
The contents of my mrbayes block is:
lset nst=2 rates=invgamma;
mcmc ngen=1000000 samplefreq=1000 printfreq=1000 temp=0.2 nruns=3 nchains=8;
sumt burnin=10000;
sump burnin=10000;
I execute it via:
$ MPI_OPTIONS="-np 24" mb-mpi
The output doesn't seem to indicate anything relevant to the problem, but it does inform me that one chain is running per MPI processor.
Has anyone encountered something similar? Am I missing something?
Assuming you are just using MPI as a mechanism to run multiple copies of the program (no internal synchronizations), you can still hurt performance of each instance. By running 24 copies on the same processor chip, you could be exhausting some system resources, such as thrashing shared caches or saturating memory bandwidth. Also, if the program is reading a large input file, having 24 processes reading the same file concurrently is less performant than having just one process reading the file.
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running