Reliable Performance Auto-Tuning in Presence of DVFS
Authors: Md Rakib Hasan (Louisiana State University), Eric Van Hensbergen (ARM Research Labs, LLC), Wade Walker (ARM Research Labs, LLC)
Abstract: In an era where exascale systems are imminent, maintaining a power budget for such systems is one of the most critical problem to overcome. Along with much research on balancing performance and power, Dynamic Voltage and Frequency Scaling (DVFS) is being used extensively to save idle-time CPU power consumption. The drawback is that the inherent random behavior of DVFS makes walltime unreliable to be used as a performance metric which causes random performance from libraries (e.g. ATLAS) that rely on machine-specific auto-tuning of several characteristics for the best performance. In this poster: 1) We show that a suboptimal selection (not the worst case) of kernel and block size during auto-tuning can cause ATLAS to lose 40% of DGEMM performance and 2) We present a more reliable performance metric in presence of DVFS that can estimate the same performance as no-DVFS yielding proper autotuning.
Two-page extended abstract: pdf