Two pervasive trends in high-performance computing, GPU acceleration and cloud computing, are increasingly being seen as a suitable match, particularly in some early-adopter industries like the life sciences.
To frame a discussion about the marriage of these two developments in HPC, Cycle Computing’s Jason Stowe recently revealed some of the results of benchmarking efforts to test how GPU-flaunting public clouds perform against in-house hardware.
Cycle Computing set about their benchmarking effort to measure the impact of virtualization on the compute performance of the GPU, the bandwidth going to the GPU and the performance of specific applications, in this case molecular dynamics (MD).
They compared results with non-virtualized physical hardware versus what they could achieve using Amazon EC2 using GPU acceleration, both of which employed NVIDIA 2050-class GPUs.
The team used the Scalable Heterogeneous Benchmarking Suite (SHOC), which was created at Oak Ridge National Laboratory and found that for the MD application in-house hardware and cloud hardware performed the same and interestingly (if not rather surprisingly) the only differences were “due to bandwidth performance characteristics that favored the cloud GPU instances.” Other than this, the maximum FLOPS were equivalent.
As Stowe noted of the results, “Cloud and GPUs are important trends in HPC, as they both offer faster time to results than alternatives… The SHOC benchmark results show that individual node performance for GPUs in the cloud and those deployed internally are comparable, besides slight differences in the bandwidth to and from the GPUs.”