Fine-grained Application Energy Modeling

Project Details

Project Lead
Catherine Olschanowsky 
Project Manager
Catherine Olschanowsky 
Supporting Experts
 
Institution
UCSD, SDSC/PMaC  
Discipline
Computer Science (401) 

Abstract

As with performance, energy-efficiency is not an attribute of a compute resource alone; it is a function of a resource-workload combination. The operation mix and locality characteristics of the applications in the workload affect the energy consumption of the resource. Data locality is the primary source of variation in energy requirements. The major contributions of this work include a method for performing fine-grained DC power measurements on HPC resources, a benchmark infrastructure that exercises specific portions of the node in order to characterize operation energy costs, and a method of combining application information with independent energy measurements in order to estimate the energy requirements for specific application-resource pairings.

Intellectual Merit

This work represents a step toward being able to better evaluate the energy-efficiency of HPC compute resource prior to acquisition. This is the fist time, to the the authors knowledge, that the energy-efficiency of compute resources can be modeled for specific applications.

Broader Impacts

Any entity that is responsible for HPC acquisitions has to take energy budgeting into account. Developing methods and tools for estimating energy costs and design trade-offs has potential uses across several government agencies.

Scale of Use

Between 2 and 4 nodes of Sierra will need to be reserved during measurements. The measurements take typically between 1 and 3 weeks.

Results

The following success story illustrates bare-metal access to FutureGrid where the user’s experiment required physical access to a FutureGrid node to attach a device needed to gather data for their research.

As with performance, energy-efficiency is not an attribute of a compute resource alone; it is a function of a resource-workload combination. The oper

ation mix and locality characteristics of the applications in the workload affect the energy consumption of the resource. Data locality is the primary source of variation in energy requirements. The major contributions of this work include a method for performing fine-grained DC power measurements on HPC resources, a benchmark infrastructure that exercises specific portions of the node in order to characterize operation energy costs, and a method of combining application information with independent energy measurements in order to estimate the energy requirements for specific application-resource pairings.

During August 2010, UCSD allocated a single node of the Sierra cluster to Olschanowsky for two weeks.  During that time Olschanowsky attached a custom-made power monitoring harness to the node as shown in Figure 1. 

Figure 1:  Attached a custom-made power monitoring harness to the node.

Fine-grained power measurements of components were taken by measuring the current close to each component; this is done using a custom harness to intercept the DC signals. Both CPUs and each memory DIMM were measured this way. The CPUs are measured by intercepting the signal at the power socket between the power supply and the motherboard; the DIMMs are measured using extender cards. In addition to the DC measurements course-grained power measurements are taken using a WattsUp device (a readily available power analyzer). Once installed a series of benchmarks were run to gather needed data for their models.   This data will be included as part of Olschanowky’s PhD dissertation.  Olschanowsky is a PhD candidate for the Department of Computer Science and Engineering at UC San Diego. 

The node that Olschanowsky used was tested and returned to service and will be recertified by IBM.