Exploring Instance Heterogeneity In Public Cloud Providers For Hpc Applications

Public cloud providers offer a wide range of instance types with different speeds, configurations, and prices, which allows users to choose the most appropriate configurations for their applications. When executing parallel applications that require multiple instances to execute, such as large scientific applications, most users pick an instance type that fits their overall needs best, and then create a cluster of interconnected instances of the same type. However, the tasks of a parallel application often have different demands in terms of performance and memory usage. This difference in demands can be exploited by selecting multiple instance types that are adapted to the demands of the application. This way, the combination of public cloud heterogeneity and application heterogeneity can be exploited in order to reduce the execution cost without significant performance loss. In this paper we conduct an evaluation of three major public cloud providers: Microsoft, Amazon, and Google, comparing their suitability for heterogeneous execution. Results show that Azure is the most suitable of the three providers, with cost efficiency gains of up to 50% compared to homogeneous execution, while maintaining the same performance.