Performance on JUWELS-Booster

[goord]

I have completed my tests on juwels-booster, which has 2 x 24C AMD-EPYC Rome cpu's and 4 x NVIDIA A100 per node. I see virtually no speedup yet between CPU/GPU versions

case	config	longwave time [ms]	shortwave time [ms]
allsky	gcc	251	257
allsky	gcc+cuda	255	258
rfmip	gcc	316	255
rfmip	gcc+cuda	323	260

For the CUDA runs I used the cuda branch.

[Chiil]

Can you also test the rcemip testcase and increase the work load to by setting n_col to 64**2? The allsky and rfmip cases are relatively cheap and are mainly used for correctness checking.

[goord]

Yes for some reason I couldn't get the input file for the rcemip case... will try once more.

[Chiil]

Send me a message on Slack if I need to clarify something, maybe the make_links.sh misses a step.

[isazi]

I managed to run all three tests on the DAS.

[goord]

Ah I first needed to run test_rcemip_input.py. Ok, here is the updated table:

case	config	longwave time [ms]	shortwave time [ms]
allsky	gcc	251	255
allsky	gcc+cuda	105	94
rfmip	gcc	316	255
rfmip	gcc+cuda	98	78
rcemip	gcc	38418	37622
rcemip	gcc+cuda	3708	3337

This is running without the 'cloud optics' flag btw.

[isazi]

Would it make sense if I also provide the same table for DAS-5? We have a node with an A100 (but older and slower Xeon CPUs).

[goord]

You can tell me if those numbers are in line with mine Alessio.

[goord]

Fixed my latest table which was transposed

[Chiil]

I did a quick benchmark on an AWS P3 V100 instance:

case	config	longwave (ms)	shortwave (ms)
rcemip	gcc+cuda	3117	2331
rcemip	gcc	14365	11362

[goord]

After using fortran compiler flags (!) I get following figures:

case	config	longwave (ms)	shortwave (ms)
rcemip	gcc+cuda	3688	3319
rcemip	gcc	8084	6982

So A100's (at least the ones in juwels-booster) appear somewhat slower than the V100 card. This may indicate the code needs some re-tuning...

[Chiil]

We have not done much tuning yet. We have rushed a little in getting a reference implementation ready that gives identical results with the CPU version, but there is probably still a lot to gain in kernel tuning.

[goord]

Some ad-hoc tuning of the kernel block sizes reduces the timings to 3051 ms (lw) and 2669 ms (sw) so yes, there is definitely headroom for speedup by tuning on A100s.

[goord]

Further manual tuning established

case	config	longwave (ms)	shortwave (ms)
rcemip	gcc+cuda	1439	1174