Way to use qbatch efficently on pbs torque ?

[rajiv03]

I've huge size(~300000000 million) of list and wants to use a java written algorithm(one executable commnd line) to do some calculation. This algorithm completes the calculation mostly takes less than a min on avg of 1-2 min.

I have been using gnu parallel but jobs are not faster/low cpu utilization. I want to use 5 nodes(each have 72 cpus) or more with pbs torque scheduler in qbatch.

fyi : gnu parallel commands I use

export JOBS_PER_NODE=72 job_parallel="parallel -j $JOBS_PER_NODE --slf $PBS_NODEFILE --wd $WDIR --joblog process.log --resume --progress"

$job_parallel -a input_list.lst sh run_input.sh {}

I would be gratful if you could suggest how do i run effiencely utiizing all the cpus for this large list? Thank you

[gdevenyi]

Hi @rajiv03

It sounds like what you need to do is overload the jobs to get high CPU utilization

First, dump the entire set of commands into a file called "joblist" or similar.

Make sure you've set up your QBATCH_* environment variables to define QBATCH_NODES, and QBATCH_PPJ=72 and QBATCH_SYSTEM="pbs"

Then I would do something like

qbatch -c (length of joblist)/5 -j (72*N) joblist

What we need to decide on is how many multiples of 72 i.e. the N value will saturate the CPU. One thing to consider however, is, is your code inefficient in some way? Perhaps you're IO limited and that's why you're not getting much speedup?

[gdevenyi]

Closing as there has been no further response from reporter.