Monitoring and Analysis of Power Consumption on HPC Clusters using XDMoD
Authors:
Open Access
- 26 July 2020
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 112-119
Abstract
As part of the NSF funded XMS project we are developing tools and techniques for the audit and analysis of HPC infrastructure. This includes a suite of tools for the analysis of HPC jobs based on performance metrics collected from compute nodes. Although it may not be salient to the user, the energy consumption of an HPC system is an important part of the cost of maintenance and contributes a substantial fraction of the cost of calculations done with the system. We added support for energy usage analysis to the open-source XDMoD tool chain. This allows HPC centers to provide information directly to HPC stakeholders about the power consumption. This includes providing end users with energy usage information about their jobs as well as providing data to allow HPC center staff to analyze how the energy usage of the system is related to other system parameters. We explain how energy metrics were added to XDMoD and describe the issues we overcame in instrumenting a 1400 node academic HPC cluster. We present an analysis of 14 months of data collected on real jobs on the cluster. We performed a machine learning analysis of the data and show how energy usage is related to other system performance metrics.Keywords
Funding Information
- National Science Foundation (ACI 1445806)
This publication has 17 references indexed in Scilit:
- A Validation of DRAM RAPL Power Measurements2016
- Data Center Energy Consumption Modeling: A SurveyIEEE Communications Surveys & Tutorials, 2015
- Extending LDMS to Enable Performance Monitoring in Multi-core Applications2015
- Open XDMoD: A Tool for the Comprehensive Management of High-Performance Computing ResourcesComputing in Science & Engineering, 2015
- Comprehensive Resource Use Monitoring for HPC Systems with TACC Stats2014
- A survey on techniques for improving the energy efficiency of large-scale distributed systemsACM Computing Surveys, 2014
- Energy Accounting and Control with SLURM Resource and Job Management System2014
- Power measurement techniques on standard compute nodes: A quantitative comparison2013
- Performance metrics and auditing framework using application kernels for high‐performance computer systemsConcurrency and Computation: Practice and Experience, 2012
- The ganglia distributed monitoring system: design, implementation, and experienceParallel Computing, 2004