High-Performance Hypervisor Architectures: Virtualization in HPC Systems

Abstract

Virtualization presents both challenges and opportunities for HPC systems and applications. This paper reviews them and also offers technical insights into two key challenges faced by future highend machines: (1) the I/O challenges they face and (2) the multicore nature of future HPC platforms. Concerning the latter, we argue the need to better structure virtualization solutions, to improve the compute experience of future HPC applications, and to provide scalability for virtualization solutions on future many-core platforms. Concerning (1), we present new methods for device virtualization, along with evidence that such methods can improve the I/O performance experienced by guest operating systems and applications running on HPC machines. Experimental results validate the claims made in (1) and (2), attained with the Xen hypervisor running on modern multi-core machines. Finally, the paper offers new directions for research in virtualization for HPC machines. It describes ways for further improving the effective I/O performance as seen by HPC applications by extending virtualization to associate application-relevant functions with I/O data streams and through QoS support for these streams. Other issues discussed are reliability enhancements for I/O, and online system and resource management.