The project examines novel services built on top of public cloud infrastructure to enable cost-effective high-performance computing. The PIs will explore the on-demand, elastic, and configurable features of cloud computing to complement the traditional supercomputer/cluster platforms. More specifically, this research aims to assess the efficacy of building dynamic cloud-based clusters leveraging the configurability and tiered pricing model of cloud instances. The scientific value of this proposal lies in the novel use of untapped features of cloud computing for HPC and the strategic adoption of small, cloud-based clusters for the purpose of developing/tuning applications for large supercomputers.
Through this research, the PIs expect to answer key research questions regarding: (1) automatic workload-specific cloud cluster configuration, (2) cost-aware and contention-aware data and task co-scheduling, and (3) adaptive, integrated cloud instance provisioning and job scheduling, plus workload aggregation for cloud instance rental cost reduction. If successful, this research will result in tools that adaptively aggregate, configure, and re-configure cloud resources for different HPC needs, with the purpose of offering low-cost R&D environments for scalable parallel applications.
