MetUM performance, ARCHER2


I am in the process of completing a Technical Assessment form for ARCHER2 usage on a (non-NERC) grant proposal.
For this, I require information on how well the UM scales on ARCHER2.
The jobs I am proposing are high resolution regional simulations.
Specifically, I need to provide a table showing how the runtime of my most computationally expensive job (I suppose the forecast jobs for my highest resolution nest) varies with the number of nodes used.
I don’t suppose you have any relevant data at hand (beyond what is available here

For what it’s worth, my high resolution jobs will be 4-km grid spacing, 70 vertical levels, 100-second timestep,

Thanks, Andy


We have used the following on occasion:

For N512, we get about 250,000 AU/model month. That’s for 66.8 million
grid points and a 600 sec time step

So, for limited area models, use 0.0037 AU/grid point and scale for
the short time step, we used 150 sec for 12k and 100 sec for 4km

for 12km use 0.015 AU/grid point
for 4km use 0.022 AU/grid point.

Roughly speaking.

But we’ve not done the same for ARCHER2 - these numbers are probably good enough to estimate CUs - I’m not sure how ARCHER will view this level of detail.


Thanks Grenville.
In fact you sent me this information a while ago, and I have used these figures to estimate the total cost of the simulations I propose.
However, this is insufficient - ARCHER2 got back to me and said they require scaling information - specifically how the runtime for a UM job varies with the number of nodes it is run on.
I don’t currently have an ARCHER2 account (I use MONSooN) so can’t test this myself. So was wondering if you might have any data on this already. Perhaps not?
Thanks, Andy

You could provide scaling info from Monsoon - it’s pretty much the same as ARCHER and I believe they accept ARCHER scaling.
The only scaling plots we have are at