Based on the benchmark below, this application does not scale beyond a single Biowulf node (FDR Infiniband, blue line in the plot below). There is no benefit to running on Infinband, therefore. (The improvement seen with FDR IB over DDR IB is due to the faster processors on the FDR IB nodes). It is best, therefore, to run Amber jobs on the non-Infiniband nodes.

If you wish to run Amber jobs on Infiniband, please run your own benchmarks to ensure that there is a benefit to using IB. (and please, send them to staff@hpc.nih.gov) Factor IX benchmark from the Amber Benchmark suite.

# cores	ns/day
	8 x 2.8 GHz Intel E5462 DDR IB	16 x 2.6 GHz Intel E5-2650v2 FDR IB
1	0.11	0.2
2	0.21	0.38
4	0.39	0.72
8	0.56	1.23
16	0.85	2.05
32	1.07	2.55
64	1.26	2.76

Amber14 Benchmark Suite, available from here.

Amber performs well on a single GPU, with a significant speedup over a single Infiniband node. The results obtained on Biowulf are very similar to the ones on the Amber website. As described on the site, some of the benchmark jobs are too small to run effectively over multiple GPUs. As always, you should run benchmarks for your own jobs.

IB FDR runs were performed on: 16 x 2.6 GHz (Intel E5-2650v2), 56 Gb/s FDR Infiniband.
GPU runs were performed on:16 x 2.6 GHz (Intel E5-2650v2) CPUs, 2 x NVIDIA K20x GPUS, 56 Gb/s FDR Infiniband

Explicit Solvent Benchmarks:

Implicit Solvent Benchmarks: