high performance computing on graphics processing units: hgpu.org

Trends indicate a rapid increase in the number of cores on chip, exhibiting various types of performance and functional asymmetries present in hardware to gain scalability with balanced power vs. performance requirements. This poses new challenges in platform resource management, which are further exacerbated by the need for runtime power budgeting and by the increased dynamics in workload behavior observed in consolidated datacenter and cloudcomputing systems. This paper considers the implications of these challenges for the virtualization layer of abstraction, which is the base layer for resource management in such heterogeneous multicore platforms. Specifically, while existing and upcoming management methods routinely leverage system-level information available to the hypervisor about current and global platform state, we argue that for future systems there will be an increased necessity for additional information about applications and their needs. This ‘end-to-end’ argument leads us to propose ‘performance points’ as a general interface between the virtualization system and higher layers like the guest operating systems that run application workloads. Building on concrete examples from past work on APIs with which applications can inform systems of phase or workload changes and conversely, with which systems can indicate to applications desired changes in power consumption, performance points are shown to be an effective way to better exploit asymmetries and gain the power/performance improvements promised by heterogeneous multicore systems.