Reduce node requests to speed up model throughput

General
1

Dear CMS help desk,

I have two suites (u-cj936 and u-cj961) that have been running extremely slowly. This is essentially because of the medium resolution (HadGEM-GC3.1-MM ) and fully-coupled configuration.
They were originally copied from the CMIP6 piControl u-aq281(https://code.metoffice.gov.uk/trac/ukcmip6/wiki/runs/u-aq281).

I would like the throughput to be improved, and one way to do that I suppose is to reduce the node hours request (and walltime accordingly of course).
The current setting is 120 hours with a walltime of 1hr45min, screenshots of the node hour settings are attached.

I wonder what is the best practise to achieve this goal? Specifically, how should I change these numbers to optimise the use of resource and performance.
It would be much appreciated if there is any suggestion or can you please point me to the right place.

Cheers,
Alcide
image

image
image483×626 45.4 KB

2

Dear CMS help desk,

I am still wondering if there is any solutions?

Regards,
Alcide

3

Alcide

What do you expect as the model performance - MO configurations are usually quick?
Throughput speed depends on a lot of factors, mostly out of your control, such as how busy the machine is and how much you have run recently. I can’t really make any useful suggestions.

Grenville

4

Hi Grenville,

Many thanks for getting back to me.

Good to know that the MO configurations are the optimal solution. However, just to clarify, these two runs are part of 10 ensemble members. So I was wondering if there is any advice on whether changing node usage per run would potentially help to speed up. Fine if not, but just in case,

Cheers, Alcide

5

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.