CPU multicore and NUMA - updated results from old GBM-multicore repo

[szilard]

Redoing old stuff from: https://github.com/szilard/GBM-multicore

New code here: https://github.com/szilard/GBM-perf/tree/master/analysis/multicore

[szilard]

Results:

https://htmlpreview.github.io/?https://github.com/szilard/GBM-perf/blob/master/analysis/multicore/results/res.html

[szilard]

2020-09-09 UPDATE: xgboost/lightgbm has improved in multi-core scaling / NUMA slow-down has been mitigated:

https://htmlpreview.github.io/?https://github.com/szilard/GBM-perf/blob/master/analysis/multicore/results-update2020sept/res-new.html

compare vs:

https://htmlpreview.github.io/?https://github.com/szilard/GBM-perf/blob/master/analysis/multicore/results-update2020sept/res-old.html

e.g.:

NUMA issue:

old:

new:

multicore scaling:

old:

new:

[szilard]

Re-run with all tools (+h2o, ++catboost):

https://htmlpreview.github.io/?https://github.com/szilard/GBM-perf/blob/master/analysis/multicore/results-update2020sept-2/res.html

Same with results for 1,2 cores removed and then rescaled to better see what's going on for many cores:

Speedups for 2,4,8,16 physical cores (no HT and no NUMA):

Speedup from 1 to 16 cores is:

size	h2o	xgboost	lightgbm	catboost
0.1M	3	6.5	1.5	3.5
1M	8	6.5	4	6
10M	24	5	7.5	8

[szilard]

runtime/size:

[szilard]

AUC:

[szilard]

NUMA+HT effect (combined):

64 cores (2 sockets, each with 16 physical cores + 16 HT) vs 16 physical cores on 1 socket:

Below red line means on 64 cores it is slower than on 16 cores

size	h2o	xgboost	lightgbm	catboost
0.1M	-40%	-50%	-70%	15%
1M	-15 %	-2%	-60%	-20%
10M	25%	35%	-20%	10%