Cloud HPC 2018-12: Google Cloud
Overview
As a natural continuation of our prior work on benchmarking cloud platforms for high-performance computing (HPC) applications in materials modeling, we decided to understand the suitability of Google Compute Engine (GCE) hardware for this purpose. In order to do so, we ran High-Performance Linpack (HPL [1]), network latency and bandwidth benchmarks [2] and Vienna Ab-initio Simulation Package (VASP [3]) for GCE.
Results
Below we demonstrate some of the results. GCE shows a rather slow interconnect network and base CPU frequency (Haswell, 2.3 GHz). Although the official documentation [4] mentions a ~20% higher turbo-frequency we were unable to achieve it while attempting the benchmarks.
Readers can find the full explanation elsewhere online [5].
Speedup Ratio
Normalized ratio of the performance for a given number of nodes to the performance for a single node.
AZ — Microsoft Azure, OL — Oracle Cloud, AWS — Amazon Web Services, GCE — Google Compute Engine.
Performance Per Core
AZ — Microsoft Azure, OL — Oracle Cloud, AWS — Amazon Web Services, GCE — Google Compute Engine.
Network Bandwidth
AZ — Microsoft Azure, OL — Oracle Cloud, AWS — Amazon Web Services, GCE — Google Compute Engine.
Conclusion
With more and more of the latest hardware coming to the cloud, it is natural to ask about its performance for specific applications. From what we are able to grasp from this quick study, running distributed memory calculations on GCE might currently present a challenge.
Links
[1] High-Performance Linpack, the official website
[2] MPI Benchmarks, Documentation
[3] Vienna Ab-initio Simulation Package, the official website
[4] Google Compute Engine CPU platforms, Documentation
[5] Exabyte.io documentation: benchmarking cloud vendors in 2018