On the Design of a Deep Computing Service Cloud

This paper presents a detailed view of a system called Deep Cloud, which enables the offering of High Performance Computing (HPC) capabilities as a service, and the design decisions and architectural considerations involved in building such a system. Our system straddles traditional Grid computing and emerging Infrastructure-as-a-service (IaaS) systems. Similar to Grid computing, our goal is to enable scalable scientific applications. However, we extend the key principles from IaaS to enable predictable resource allocation of HPC resources with a pay-as-you-go model. Our approach uses dynamic pricing to shape demand for HPC resources, while providing access to HPC resources for both interactive and batched scientific workloads. We also describe a novel approach to HPC resource catalog creation: the conversion of an HPC 3D topology into an easily calculable resource inventory that supports predictable reservation and scheduling. This enables a dynamic resource placement mechanism for reservation requests which we call Just-In-Time placement. Just-In-Time placement collectively performs resource placement to achieve optimal HPC resource utilization. Finally, we describe the need for a coherent, integrated user experience when interacting with the system.