Scalable System-level Active Low-Power Mode with Bounded Latency

Many system-level inactive low power modes exploit idle periods to obtain energy savings. With the emergence of multicore servers, idle periods are becoming increasingly rare. In order to save energy in multicore servers, low-utilization periods, which remains with increasing core count, must be exploited. Server-level heterogenous servers, such as KnightShift, have been shown to significantly improve the energy proportionality of datacenter servers through exploiting lowutilization periods. However, previous switching policies, which decides when to switch between a high-power highperformance node and a low-power lower-performance node, are simplistic and easily fooled by server utilization patterns, leading to false switches and thrashing causing unbounded latency impact. In this paper, we propose Hueristic-based Switching Policies (HSP), which uses utilization history to predict when future high utilization periods will occur. We show that HSP can significantly reduce thrashing and false switches, bounding latency while still maintaining significant energy savings. Furthermore, we show that active low-power modes that exploit low utilization periods are able to sustain energylatency tradeoffs as core count increases and offer superior energy savings compared to idleness scheduling algorithms.