MPI-like Framework for Cloud

About

Introduction: Message Passing Interface (MPI) has been the predominant standardized system for writing parallel and distributed applications. However, MPI is not the system of choice for cloud platforms due to the lack of low latency high bandwidth network capabilities and different computing infrastructure. Studies have shown that the performance of an application, when ported to the cloud, could be 35x slower if using state-of-the-art MPI implementations. Creating an MPI-like framework for cloud platforms is thus an unmet need to improve a software developer’s productivity and reduce the complexity of an HPC application. Technology: Georgia State University researchers have developed a cloudMPI framework for writing new MPI-style applications and porting MPI-based applications to the Azure cloud platform. The researchers identified the pain points for designing and implementing an MPI-like framework for cloud platforms so this framework could benefit eScience developers to write MPT-style applications for the Azure cloud platform. Results demonstrate that the performance of a complex application, Crayons, is comparable using cloudMPI and using native Azure APIs in terms of network traffic, number of tenants on a physical machine, and distance between compute instances and data.