Open aydinsaribudak opened 4 years ago
Hi, I would like to share our assumption for expected allocation, if we ran weak-scaling experiments on Bridges. I will be glad to hear your suggestions.
3 experiments will be conducted on K80 and P100 nodes for each md_count. The allocations are as follows.
For 60 md_counts: 3 exp 15 nodes 4 gpu 12 hours 1 SU/gpu-hour = 2,160 SUs. For 120 md_counts: 3 exp 12 hours (16 nodes 4 gpu 1 SU/gpu-hour + 28 nodes 2 gpu 2.5 SU/gpu-hour) = 7,344 SUs. For 240 md_counts: 3 exp 12 hours (16 nodes 4 gpu 1 SU/gpu-hour 2 times + 32 nodes 2 gpu 2.5 SU/gpu-hour + 24 nodes 2 gpu 2.5 SU/gpu-hour ) = 14,688 SUs. For 480 md_counts: 3 exp 12 hours (16 nodes 4 gpu 1 SU/gpu-hour 4 times + 32 nodes 2 gpu 2.5 SU/gpu-hour 3 times + 16 nodes 2 gpu * 2.5 SU/gpu-hour ) = 29,376 SUs.
53,568 SUs in total
Thanks, Aydin
Hi, I am planning to run a set of experiments on Summit as described below. I am creating this ticket to make an estimation for the expected allocation that I will use on Summit to conduct these experiments. 3 experiments: 10 nodes, wall time is 12 hours -> 360 node-hours 3 experiments: 20 nodes, wall time is 12 hours -> 720 node-hours 3 experiments: 40 nodes, wall time is 12 hours -> 1440 node-hours 3 experiments: 80 nodes, wall time is 12 hours -> 2880 node-hours in total, 5400 node-hours. The goal of these experiments is to compute total execution time and overhead for the MD simulations (for adrp protein folding). Thanks.