Opened 7 years ago

Closed 7 years ago

#1049 closed help (fixed)

Resource estimate

Reported by: simon.tett Owned by: um_support
Component: UM Model Keywords:
Cc: Platform: HECToR
UM Version: <select version>

Description

Hi,

I've got a potential new colleagues coming to join the University of Edinburgh as a lecturer. This colleague wants to do some simulations — and I'd like to provide some guidance to him. This is an urgent request as I'd like to complete the negotiations with the candidate very soon. I'm not after very exact estimates — more a sense of what would be needed. If the resource requirements are not too large we may be able to allocate some resource from the share of HECToR that is available to UoE researchers. Otherwise I'll suggest he needs to write a proposal for HECToR time.

The candidate writes "* Computational cores: I have made a tentative estimate of the total resources based on running an atmospheric and a coupled model (at times, simultaneously) at about 2-degree resolution for a number of experiments running each for several decades of simulations, and using multiple ensembles (ideally also including full chemistry). The targeted output would be of 4-5 years of simulation per day."

My suggestion to him would be to use use UKCA on HECToR so my questions are:
1) Is this a sensible recommendation?
2) How much resource (HECToR units) per simulated year are required for both ocean/atmosphere/chemistry configuration and how much for atmosphere/chemistry configuration.
3) How many simulated years/real-time-day would be feasible?
4) How many cores would be needed?

Simon

Change History (3)

comment:1 Changed 7 years ago by luke

Hi Simon

I can't comment on how your colleague would get HECToR time, and how much would be available. The advice I have had in the past is that some time can be made available for short-ish runs and testing, but longer runs usually need to have time applied for specifically. However, this is often partly dependent on how much time is left in the pot!

UKCA on HECToR runs in the 4 node (128 core) queue for N48L60 jobs (3.75x2.5 up to 84km) and the 8 node (256 core) queue for N96L63 jobs (1.875x1.25 up to 41km). These are for UM 7.3 HadGEM3-A r2.0 (equivalent to "Global Atmosphere 2.0" in new money I think). UM8.2+ jobs should be coming online soonish at N96L85 (L63 troposphere + L60 stratosphere up to 85km) and this model will have some optimisations, but I'm not sure how it will run or scale on HECToR yet.

It is possible to run the atmosphere model very quickly on HECToR, easily getting up and over the 4-5 model years per day. UKCA will slow the model down by a factor of 3-6, depending on the chemistry scheme used. The N96L63 tropospheric chemistry might only get up to 12 months in one day, whereas the N48L60 running stratospheric chemistry could get up to 3 years a day, assuming that these ran contently in the queue. The stratosphere-troposphere scheme will run about ½ as fast, so about 18 months a day.

Running a coupled model is a bit more complicated with the way the coupler and ocean work with the architecture of the machine, so will be slower again.

From the HECToR AU calculator at http://www.hector.ac.uk/howcan/admin/apply/calculator/ I get 30 years of N48L60 to be 224kAUs a will take at least 10 days to run, although with queuing it will be longer than this, and what exactly the number kAUs mean in terms of available resources I can't comment on.

Without knowing more about the type of chemistry required or the processes being studied I can't comment further. Using UKCA on HECToR could certainly be suitable though.

Thanks,

Luke

comment:2 Changed 7 years ago by simon.tett

Hi Luke,

thanks very much for that. I think if candidate accepts the job the University of Edinburgh will provide him with time from our Allocation (10%) it gets as part of the agreement for running HECToR.

Simon

comment:3 Changed 7 years ago by grenville

  • Resolution set to fixed
  • Status changed from new to closed
Note: See TracTickets for help on using tickets.