Evaluating HPC-Style CPU Performance and Cost in Virtualized Cloud Infrastructures
Jay Tharwani, Shobhit Aggarwal, Arnab A Purkayastha
TL;DR
The study addresses variability in CPU performance and cost for HPC-style workloads across major cloud providers and CPU architectures. It benchmarks a subset of SPEC ACCEL OpenMP workloads on Intel, AMD, and ARM instances from AWS, Azure, GCP, and OCI under on-demand and one-year pricing, using standardized configurations to isolate CPU effects. Key findings show AWS delivers the fastest runtimes but at a premium, while OCI offers the most cost-effective option with slower performance; Azure and GCP fall in between, with GCP ARM being notably slower and more expensive. The work provides practical guidance for balancing performance and cost in cloud HPC, and outlines avenues for future research in multi-node scaling, specialized HPC instances, and alternative pricing models.
Abstract
This paper evaluates HPC-style CPU performance and cost in virtualized cloud infrastructures using a subset of OpenMP workloads in the SPEC ACCEL suite. Four major cloud providers by market share AWS, Azure, Google Cloud Platform (GCP), and Oracle Cloud Infrastructure (OCI) are compared across Intel, AMD, and ARM general purpose instance types under both on-demand and one-year discounted pricing. AWS consistently delivers the shortest runtime in all three instance types, yet charges a premium, especially for on-demand usage. OCI emerges as the most economical option across all CPU families, although it generally runs workloads more slowly than AWS. Azure often exhibits mid-range performance and cost, while GCP presents a mixed profile: it sees a notable boost when moving from Intel to AMD. On the other hand, its ARM instance is more than twice as slow as its own AMD offering and remains significantly more expensive. AWS's internal comparisons reveal that its ARM instance can outperform its Intel and AMD siblings by up to 49 percent in runtime. These findings highlight how instance choices and provider selection can yield substantial variations in both runtime and price, indicating that workload priorities, whether raw speed or cost minimization, should guide decisions on instance types.