Adaptive Power Shifting for Power-Constrained Heterogeneous Systems

The number and heterogeneity of compute devices, even within a single node, has been steadily on the rise. Since all systems must operate under power cap, discrete devices that can run simultaneously at their highest frequency is limited by globally-imposed cap. Current incorporate centralized management unit statically controls distribution among node. However, such static policies are unaware dynamic utilization profile across which leads to unfair allocations end up degrading system throughput performance. problem particularly acute in presence since type-specific performance-boost capabilities cannot be leveraged via utilization-agnostic allocations. This paper proposes Adaptive Power Shifting for multi-accelerator heterogeneous (APS), technique leverages information dynamically allocate re-distribute budgets multiple devices. Democratizing allocation based needs results dramatic speedup over need-agnostic allocation. We use APS real OpenPOWER node with 2 CPUs 4 GPUs demonstrate value on-demand, equitable Overall, proposed solution increases performance respect two state-of-the-art techniques 14.9% 13.8%.