MPI Latency on Google Compute Engine

Google has entered the Infrastructure as a Service (IaaS) cloud computing market with its Google Compute Engine (GCE) service, emphasizing its networking infrastructure performance. A test was conducted to evaluate the latency within a high-performance computing (HPC) application context, specifically measuring the latency between machines in an MPI cluster using the osu_latency benchmark. The results showed similar latency numbers across different GCE instance types, with some variance likely due to machine contention. Comparatively, Amazon EC2 instances demonstrated lower latencies, but this comparison was not entirely equivalent as Amazon offers specialized cluster compute instances and placement groups that enhance performance. The findings suggest GCE may currently be better suited for workloads that are more parallel in nature, while highlighting that micro benchmarks may not fully represent real-world application performance. Future improvements in HPC performance are anticipated as competition among cloud providers like Google, Amazon, and Microsoft intensifies, benefiting consumers by driving advancements and options in cloud computing services.