Manila: Using a Densely Populated PMC-Space for Power Modelling within Large-Scale Systems

摘要

In this paper, we propose a new power modelling method, called Manila, that can largely reduce the effort of PMC-based power modelling using high-dimensional k-nearest neighbour searching, without the use of model tuning and domain-specific knowledge. This method helps improve the accuracy of PMC-based power modelling and widen its scope of use. Specifically, Manila uses a parameterised micro-benchmark to automatically generate a densely populated PMC-space that represents a large variety of computing workloads, which is essential for increasing the accuracy of power modelling and widening its scope of use. This is in contrast to current PMC-based power models, that have a sparse PMC-space, due to using predefined benchmarks. Since the micro-benchmark is independent from any applications and can generate the generic computing workloads of many applications, our method is more widely extendable and applicable than the existing methods. Manila can efficiently search the dense PMC-space for power estimates using a nearest neighbour search algorithm. Experimental results demonstrate that Manila is more accurate in power measurements for a wide range of parallel benchmarks, with a mean absolute error of 2.8%.