A component infrastructure for performance and power modeling of parallel scientific applications

Abstract

Characterizing the performance of scientific applications is essential for effective code optimization, both by compilers and by high-level adaptive numerical algorithms. While maximizing power efficiency is becoming increasingly important in current high-performance architectures, there is little or no hardware or software support for detailed power measurements. Hardware counter-based power models are a promising method for guiding software-based techniques for reducing power. We present a component-based infrastructure for performance and power modeling of parallel scientific applications. The power model leverages on-chip performance hardware counters and is designed to model power consumption for modern multiprocessor/multicore systems. Our tool infrastructure includes application components as well as performance/power measurement and analysis components. We collect performance data using the TAU performance component and apply the power model in the performance and power analysis of a PETSc-based parallel fluid dynamics application by using the PerfExplorer component.

Understanding hardware and software metrics with respect to power consumption

“…This benchmark is denoted as idl. 2. linpack: This pre-compiled linear algebra code from Intel contains the optimized LINPACK [31] benchmark 5 . Internally using MKL libraries, it performs FPU/ALU instructions with the purpose of utilizing the CPU.…”

“…However, acquisition costs and deployment of power measurement devices can be, due to the nature of the platforms and number of nodes, infeasible. Recent research has significantly demonstrated that a promising alternative in order to mitigate this issue is the design of power models [3,4,5]. Taking into account that most of the current processors feature a large set of hardware counters, temperature sensors, and resource usage statistics provided by the operating system, one could cleverly use this information to predict power drawn by individual components and system power consumption.…”

“…As highlighted in [7], a good power model should always be accurate, simple, inexpensive and portable. Indeed, recent works have shown that many power models can fulfill all these properties while providing fairly good estimations [3,4,5]. In this sense, a classical methodology to validate the accuracy of power models has only been carried out by calculating the preciseness and responsiveness of their predictions [8].…”

TProf: An energy profiler for task-parallel programs

“…The regression-based power modeling approach used in TProf to account energy between concurrently executing tasks follows a long line of earlier work that derives power models using input from hardware performance counters [4,8,9,10,23,36]. Prior work establishes power models composed of linear or piecewise polynomial functions of sampled hardware event rates, using events that correlate strongly with power consumption.…”

An automated component-based performance experiment environment

“…In the database component, the user can set parameters such as the name of the configuration and the name of the loader utility in cases where an external, non component tool is exploited to import the data into a database. Other implementations of the database interface do not rely on external tool support for importing data into the database, for example the database components available as part of CQoS infrastructure [6]. This design allows the easy addition of new components to support multiple data formats, transfer protocols, and database interfaces.…”