PMaC's green queue: a framework for selecting energy optimal DVFS configurations in large scale MPI applications

Abstract

This article presents Green Queue, a production quality tracing and analysis framework for implementing application aware dynamic voltage and frequency scaling (DVFS) for message passing interface applications in high performance computing. Green Queue makes use of both intertask and intratask DVFS techniques. The intertask technique targets applications where the workload is imbalanced by reducing CPU clock frequency and therefore power draw for ranks with lighter workloads. The intratask technique targets balanced workloads where all tasks are synchronously running the same code. The strategy identifies program phases and selects the energy-optimal frequency for each by predicting power and measuring the performance responses of each phase to frequency changes. The success of these techniques is evaluated on 1024 cores on Gordon, a supercomputer at the San Diego Supercomputer Center built using Intel Xeon E5-2670 (Sandybridge) processors. Green Queue achieves up to 21% and 32% energy savings for the intratask and intertask DVFS strategies, respectively. The contributions of this work include:Detecting and characterizing program phases using interprocedural analysis on application trace data. A structural phase detection algorithm, as opposed to a time slicing method, allows phase length and location to adapt to boundaries naturally present in programs, such as loop and function boundaries. Interprocedural analysis further improves upon this because these boundaries are not limited to the scope of single loops or functions. Efficiently determining optimal frequency configurations over all phases in a program. Selecting an optimal configuration by direct measurement requires several runs for each phase to locate the optimal frequencies because the length of a phase is short compared with the time granularity needed to measure power. This work proposes a method of combining power modeling and performance measurements to predict energy consumption for frequency configurations using one application run per frequency setting. A production-quality tracing and analysis framework. Green Queue provides a robust framework for generating and analyzing application trace data. The production, organization, and processing of application trace data is automatic and largely opaque to the user, managed by a few simple command line tools. This allows for comprehensive energy management strategies to be built to leverage a wide array of detailed application statistics and for those strategies to be easily utilized. An evaluation of Green Queue on a full compute rack of Gordon. Gordon is the latest supercomputer at the San Diego Supercomputer Center built with Intel Xeon E5-2670 (Sandybridge) processors. With 1024 cores per rack, this work demonstrates that DVFS is a viable strategy for achieving significant energy savings at scale on modern hardware. Experiments demonstrating the value of application aware DVFS over blind DVFS. This work demonstrates that application aware DVFS can significantly outperform techniques that do not r...

Evaluation of DVFS techniques on modern HPC processors and accelerators for energy‐aware applications

“…14 Given the increasing interest in controlling and reducing energy dissipation, recent high-end processors have started to support advanced power-saving and power-monitoring technologies, similar to those early adopted in low-power processors. 15,16 In this paper, we perform a similar analysis on state-of-the-art HPC At variance with previous works such as previous studies, 8,10 we do not attempt to provide an automatic tuning of processors frequencies nor an automatic tool able to identify where to apply frequency scaling.…”

On low‐power SoCs as storage bricks for Bioinformatics

“…Recent strategies employ Dynamic Voltage and Frequency Scaling (DVFS) to reduce the processor frequency during I/O operations, since these do not require much processing power. DVFS has been a popular technique to save energy consumption by lowering the frequency of idle cores in multi‐core systems, idle resources in cloud environments, and during MPI applications' communication phases …”

Multi-Objective region-Aware optimization of parallel programs

“…However, recent work shows an inarguable attention to tune applications regarding several objectives. Besides wall time, energy consumption entailed by a program's execution is becoming a popular objective [29,30,31,32,33,34,35,2,36]. Resource usage [12,36], compilation time, or the size of the executable binary [37,38,39] also received attention in related work.…”