Exploiting variability for energy optimization of parallel programs

Abstract

In this paper we present optimizations that use DVFS mechanisms to reduce the total energy usage in scientific applications. Our main insight is that noise is intrinsic to large scale parallel executions and it appears whenever shared resources are contended. The presence of noise allows us to identify and manipulate any program regions amenable to DVFS. When compared to previous energy optimizations that make per core decisions using predictions of the running time, our scheme uses a qualitative approach to recognize the signature of executions amenable to DVFS. By recognizing the "shape of variability" we can optimize codes with highly dynamic behavior, which pose challenges to all existing DVFS techniques. We validate our approach using offline and online analyses for one-sided and two-sided communication paradigms. We have applied our methods to NWChem, and we show best case improvements in energy use of 12% at no loss in performance when using online optimizations running on 720 Haswell cores with one-sided communication. With NWChem on MPI two-sided and offline analysis, capturing the initialization, we find energy savings of up to 20%, with less than 1% performance cost.

Energy efficiency of nonlinear domain decomposition methods

“…Here, the baseline power of the chip P base should be the dominating factor. At constant clock speed, the total power consumption P t of a processor chip with c active cores can be approximated by (31) P t (c) = P base + cP core ,…”

“…[14]). The unknown values of P core and P base can be determined by fitting Equation (31) to power measurements on single sockets when running the Newton-Krylov-FETI-DP solver with varying the numbers of cores from 1 to 10, i.e. c=1,...,10.…”

“…Reducing energy consumption of HPC applications has gained substantial attraction in the last decade. In the context of MPI several algorithms and runtime systems have been proposed (Hsu and Feng, 2005; Kappiah et al, 2005; Kerbyson et al, 2011; Lavrijsen et al, 2016; Li et al, 2010; Lim et al, 2011; Rountree et al, 2009; Vishnu et al, 2013), which typically follow the same idea: processes exhibiting a high slack time, i.e. time they are blocked with waiting for other processes or communicating, can be clocked down to save energy.…”

Improving efficiency of semi‐direct møller–plesset second‐order perturbation methods through oversubscription on multiple nodes

“…A 2011 paper showed overall power requirements were reduced by 8% by mapping points of global synchronization within algorithms to when cores were expected to be in an idle state and manipulating the power states of the cores accordingly . In 2016, it was reported that DVFS, applied broadly to “amenable regions” of the NWChem computational chemistry code, could save up to 20% in total energy requirements with less than 1% loss in performance. Another pair of recent articles employs a similar strategy, applying DVFS to specific regions of Møller–Plesset second‐order perturbation theory (MP2) algorithms lowered overall energy use by up to 10% with minimal costs to performance.…”

Saving time and energy with oversubscription and semi‐direct M øller– P lesset second order perturbation methods

“…Through judicious application of DVFS strategies within specific phases of different Møller–Plesset Second Order Perturbation Theory (MP2) algorithms, this technique has exhibited up to a 10% overall savings in energy consumption. A 2016 report demostrated the ability of DVFS to reduce energy use up to 20% using Coupled Cluster and Density Functional Theory algorithms with as little as 1% performance degradation.…”