Virtual Accelerator Support for HPC Clouds

Due to huge levels of computing parallelism and higher performance per watt, comuting accelerators are crucial for increasing HPC efficiency. This is economically significant for small and medium size manufacturing companies (SMMs) and essential for transition to Exascale computing. However, due to high procurement costs, in-house maintenance of an accelerated HPC cluster is prohibitive for many SMMs. HPC as a service (Cloud-based computing) is the most promising alternative solution for industries to reduce their owndership costs and improve flexibility. This requires efficient virtualization and sharing of accelerators in the Cloud environments. While some support for accelerator-based HPC services exists, the existing services are inadequate, expensive, not fault-tolerant and energy efficient. This is a great barrier in front of Cloud providers to offer such services and for SMMs to them for their simulation needs. In this STTR project, RNET and ANL are proposing a virtual accelerator framework that enables Cloud-based HPC service providers to offer cost-effective, elastic and fault tolerant accelerator-based HPC services. Such product is crucial in facilitating the adoption of HPC software as a service in manufacturing and engineering industry. Commercial Applications and Other Benefits: This solution will enable HPC service providers to provide a more elastic cloud-based HPC service with fault-tolerance and power-savings for HPC applications. This will reduce the service prices and increase the overall ability of HPC services to manufacturing and industry sector, making it easier for them to utilize HPC solutions for their simulation and analysis needs.