Using multiple CPUs of a Ubutun server to run CLM5.0

[hyf412694462]

Dear all,

I have a question about how to use multiple CPUs of a Ubutun server to run CLM5.0.

The machine I used to run CLM is a Ubutun server (no batch system) with 40 Intel(R) Xeon(R) Silver 4210 CPUs. One CPU has 10 cores and 20 threads. I wonder how to set MAX_TASKS_PER_NODE and MAX_MPITASKS_PER_NODE or I need to set some other variables in config_machines.xml to utilize all cores and threads of 15 Intel CPUs to run the model? Moreover, should I change some specific settings during case.setup and case.build?

I appreciate any assistance you could provide and I look forward to hearing from you.

 

[jedwards]

I think that you would set MAX_TASKS_PER_NODE=800 and MAX_MPITASKS_PER_NODE=400 but I have not seen this machine before.

 

[hyf412694462]

Dear Jedwards,

Thank you a lot for your kind assistance. I got it. Could you please explain more about how to determine this values (400 and 800)?

 

[jedwards]

The MAX_MPITASKS_PER_NODE is usually set to the maximum number of physical cores on a shared memory node.
the MAX_TASKS_PER_NODE is set to the maximum number of allowed threads on a shared memory node. But now rereading your message I think perhaps you used CPU when you meant node - in that case the values should be MAX_TASKS_PER_NODE=20 and MAX_MPITASKS_PER_NODE=10.
And you have 15 nodes - is that correct?

 

[hyf412694462]

Dear Jedwards,

Thank you very much for you quick response. I think one CPU means one node in my case. Therefore, setting MAX_TASKS_PER_NODE=20 and MAX_MPITASKS_PER_NODE=10 is correct in my case.

However, I want to use 15 nodes to run the model, how to specify 15 (nodes)?
I would will to appreciate any help you could provide.

 

[jedwards]

You can specify NTASKS=150 or NTASKS=-15
the negative value tells cime to calculate ntasks as -1*(NTASKS)*MAX_MPITASKS_PER_NODE