Performance and Cost Analysis of AWS Burstable Instances for HPC with NAS Parallel Benchmarks

• Artur Luiz Rizzato Toru Soda UFSC
• Vanderlei Munhoz Pereira Filho UFSC / University of Bordeaux / CNRS / Bordeaux INP / Inria / LaBRI
• Márcio Castro UFSC

Resumo

Traditional High Performance Computing (HPC) typically requires expensive and complex physical infrastructures. With the advancement of cloud computing, providers such as Amazon Web Services (AWS) have begun to offer more economically viable alternatives, such as burstable and spot instances. However, selecting the appropriate resources remains challenging due to performance variability and the costs associated with different workload profiles. In this context, this work evaluates the feasibility of using burstable instances for HPC. To this end, an in-depth study was carried out using AWS on-demand burstable and non-burstable instances, evaluating their performance and cost under different HPC application scenarios. Three main sets of experiments were conducted: (1) comparative evaluation on homogeneous clusters of up to 4 nodes using NAS Parallel Benchmarks (NPB), contrasting burstable and similar non-burstable instances; (2) scalability analysis on larger clusters (4–16 nodes) with a broader variety of NPB applications; (3) isolated measurement of the impact of network bandwidth using an MPI microkernel benchmark. The results show that the ideal instance type depends on the workload profile: non-burstable instances provide the best cost-benefit for intensive and distributed workloads, while burstable standard may be more economical for light and short workloads, and burstable unlimited ensures constant performance at a higher cost for long-running workloads.

Palavras-chave: Cloud computing, Web services, High performance computing, Scalability, Conferences, Computer architecture, Bandwidth, Benchmark testing, Cost benefit analysis, Standards, high performance computing, cloud computing, NAS parallel benchmarks, burstable instances