Vishakha Gupta

GViM: GPU-accelerated virtual machines

The use of virtualization to abstract underlying hardware can aid in sharing such resources and in efficiently managing their use by high performance applications. Unfortunately, virtualization also prevents efficient access to accelerators, such as Graphics Processing Units (GPUs), that have become critical components in the design and architecture of HPC systems. Supporting General Purpose computing on GPUs (GPGPU) with accelerators from different vendors presents significant challenges due to proprietary programming models, heterogeneity, and the need to share accelerator resources between different Virtual Machines (VMs). To address this problem, this paper presents GViM, a system designed for virtualizing and managing the resources of a general purpose system accelerated by graphics processors. Using the NVIDIA GPU as an example, we discuss how such accelerators can be virtualized without additional hardware support and describe the basic extensions needed for resource management. Our evaluation with a Xen-based implementation of GViM demonstrate efficiency and flexibility in system usage coupled with only small performance penalties for the virtualized vs. non-virtualized solutions.

Kinship: efficient resource management for performance and functionally asymmetric platforms

On-chip heterogeneity has become key to balancing performance and power constraints, resulting in disparate (functionally overlapping but not equivalent) cores on a single die. Requiring developers to deal with such heterogeneity can impede adoption through increased programming effort and result in cross-platform incompatibility. We propose that systems software must evolve to dynamically accommodate heterogeneity and to automatically choose task-to-resource mappings to best use these features. We describe the kinship approach for mapping workloads to heterogeneous cores. A hypervisor-level realization of the approach on a variety of experimental heterogeneous platforms demonstrates the general applicability and utility of kinship-based scheduling, matching dynamic workloads to available resources as well as scaling with the number of processes and with different types/configurations of compute resources. Performance advantages of kinship based scheduling are evident for runs across multiple generations of heterogeneous platforms.