Computational Systems and Technology

Cloud Computing on a Grand Scale: ISI Papers Document High-Performance Feasibility

ISI’s high-performance, heterogeneous cloud group – a unit of its Computational Systems & Technology division – had two papers presented at the IEEE International Conference on Cloud Computing. Both papers show that high-performance applications can be virtualized with very little overhead, suggesting that users soon may be able to conduct supercomputer-level processing through the cloud, using disparate hardware and software of their own choosing. The IEEE conference was held in Anchorage, Alaska from June 27th to July 2nd, 2014.

The first paper shows that applications written in the OpenCL language, and that target accelerators in the latest generation of GPUs, generally can be run from virtual hosts with less than 1% overhead. Called "GPU-Passthrough Performance: A Comparison of KVM, Xen, VMWare ESXi, and LXC for CUDA and OpenCL Applications”, the paper is authored by JP Walters, David Kang, Ke-Thia Yao, Mikyung Kang, and Steve Crago of USC ISI, and Andrew Younge and Geoffrey Fox of Indiana University.

The second paper shows that SR-IOV can be used to access an Infiniband network from a virtual host with little overhead, assuming the network interface controller's parameters are tuned correctly. "Bridging the Virtualization Performance Gap for HPC using SR-IOV for InfiniBand”, written by Malek Musleh and Vijay Pai of Purdue University and JP Walters and Steve Crago of USC ISI, grew out of between Musleh and Pai while they were a visiting research assistant and visiting scholar, respectively, at ISI’s Arlington, Virginia offices.

Both papers demonstrate that users soon will be able to allocate high-performance computing resources – from both commercial cloud providers and private clouds – via software images of their own choosing and the latest heterogeneous hardware. At the same time, users will be able to capitalize on the benefits of centrally managed, dynamically allocated (“elastic”) hardware resources. ISI’s high-performance, heterogeneous cloud group is helping create that reality.
 

Groups: