Toward Implementation of a Software Defined Cloud on a Supercomputer

Abstract

Conventional cloud computing architectures may seriously constrain computational throughput for high performance computing (HPC) and high-performance data (HPD) applications. The traditional approach to circumvent such problems has been to map these applications and problems onto other specialized hardware and coprocessor architectures. This is both time and resource expensive, and poses a challenge for rapidly rising demands for computation and data analytics. In this paper we report on progress to develop an alternative experimental software defined cloud implementation that virtualizes the topology of a standard HPC computational architecture. This software defined system re-arranges access to the nodes and dynamically customizes the features of the HPC hardware architecture so that they map to the specifics of the computation and data analysis application. This allows a cloud computing implementation to utilize the specialized infrastructure capabilities of an HPC system. We have created this type of user reconfigurable architecture on an IBM Blue Gene/P supercomputing environment at the Department of Energy's Argonne Leadership Computing Facility (ALCF). This pilot configuration was implemented using both an open source cloud technology called VCL (Virtual Computing Laboratory) in combination with a provisioning module called Kittyhawk. Cloud security is addressed by configuring and running a root-less version of the VCL cloud system on the ALCF's Blue Gene/P login node.