Sarah S Powers

An Analysis Framework for Investigating the Trade-Offs between System Performance and Energy Consumption in a Heterogeneous Computing Environment

Rising costs of energy consumption and an ongoing effort for increases in computing performance are leading to a significant need for energy-efficient computing. Before systems such as supercomputers, servers, and datacenters can begin operating in an energy-efficient manner, the energy consumption and performance characteristics of the system must be analyzed. In this paper, we provide an analysis framework that will allow a system administrator to investigate the tradeoffs between system energy consumption and utility earned by a system (as a measure of system performance). We model these trade-offs as a bi-objective resource allocation problem. We use a popular multi-objective genetic algorithm to construct Pareto fronts to illustrate how different resource allocations can cause a system to consume significantly different amounts of energy and earn different amounts of utility. We demonstrate our analysis framework using real data collected from online benchmarks, and further provide a method to create larger data sets that exhibit similar heterogeneity characteristics to real data sets. This analysis framework can provide system administrators with insight to make intelligent scheduling decisions based on the energy and utility needs of their systems.

Utility Driven Dynamic Resource Management in an Oversubscribed Energy-Constrained Heterogeneous System

In this paper, we address the problem of scheduling dynamically-arriving tasks to machines in an oversubscribed heterogeneous computing environment. Each task has a monotonically decreasing utility function associated with it that represents the utility (or value) based on the tasks completion time. Our system model is designed based on the environments of interest to the Extreme Scale Systems Center at Oak Ridge National Laboratory. The goal of our scheduler is to maximize the total utility earned from task completions while satisfying an energy constraint. We design an energy-aware heuristic and compare its performance to heuristics from the literature. We also design an energy filtering technique for this environment that is used in conjunction with the heuristics. The filtering technique adapts to the energy remaining in the system and estimates a fair-share of energy that a tasks execution can consume. The filtering technique improves the performance of all the heuristics and distributes the consumption of energy throughout the day. Based on our analysis, we recommend the level of filtering to maximize the performance of scheduling techniques in an oversubscribed environment.

Falcon: Visual analysis of large, irregularly sampled, and multivariate time series data in additive manufacturing ☆ ☆☆

Abstract Flexible visual analysis of long, high-resolution, and irregularly sampled time series data from multiple sensor streams is a challenge in several domains. In the field of additive manufacturing, this capability is critical for realizing the full potential of large-scale 3D printers. In this paper, we propose a visual analytics approach that helps additive manufacturing researchers acquire a deep understanding of patterns in log and imagery data collected by 3D printers. Specific goals include discovering patterns related to defects and system performance issues, optimizing build configurations to avoid defects, and increasing production efficiency. We introduce Falcon, a new visual analytics system that allows users to interactively explore large, time-oriented data sets from multiple linked perspectives. Falcon provides overviews, detailed views, and unique segmented time series visualizations, all with adjustable scale options. To illustrate the effectiveness of Falcon at providing thorough and efficient knowledge discovery, we present a practical case study involving experts in additive manufacturing and data from a large-scale 3D printer. Although the focus of this paper is on additive manufacturing, the techniques described are applicable to the analysis of any quantitative time series.

Energy-aware resource management for computing systems

This corresponds to the material in the invited keynote presentation by H. J. Siegel, summarizing the research in [1], [2]. We address the problem of assigning dynamically-arriving tasks to machines in a heterogeneous computing environment. These machines execute a workload composed of different tasks, where the tasks have diverse computational requirements. Each task has a utility function associated with it that represents the value of completing that task, and this utility decreases the longer it takes a task to complete. The goal of our resource manager is to maximize the sum of the utilities earned by all tasks arriving in the system over a given interval of time, while satisfying an energy constraint. We describe example energy-aware resource management methods to accomplish this goal, and compare their performance. We also study the bi-objective problem of maximizing system utility and minimizing the system energy consumption. This analysis technique allows system administrators to investigate the trade-offs between these conflicting goals.

A study of the impact of scheduling parameters in heterogeneous computing environments

This paper describes a tool for exploring system scheduler parameter settings in a heterogeneous computing environment. Through the coupling of simulation and optimization techniques, this work investigates optimal scheduling intervals, the impact of job arrival prediction on scheduling, as well as how to best apply fair use policies. The developed simulation framework is quick and modular, enabling decision makers to further explore decisions in real-time regarding scheduling policies or parameter changes.

OpenSHMEM Implementation of HPCG Benchmark

We describe the effort to implement the HPCG benchmark using OpenSHMEM and MPI one-sided communication. Unlike the High Performance LINPACK (HPL) benchmark that places emphasis on large dense matrix computations, the HPCG benchmark is dominated by sparse operations such as sparse matrix-vector product, sparse matrix triangular solve, and long vector operations. The MPI one-sided implementation is developed using the one-sided OpenSHMEM implementation. Preliminary results comparing the original MPI, OpenSHMEM, and MPI one-sided implementations on an SGI cluster, Cray XK7 and Cray XC30 are presented. The results suggest the MPI, OpenSHMEM, and MPI one-sided implementations all obtain similar overall performance but the MPI one-sided implementation seems to slightly increase the run time for multigrid preconditioning in HPCG on the Cray XK7 and Cray XC30.

Using Hybrid Model OpenSHMEM + CUDA to Implement the SHOC Benchmark Suite

This work describes the process of porting the Scalable HeterOgeneous Computing (SHOC) benchmark suite from the hybrid MPI + CUDA implementation to OpenSHMEM + CUDA. SHOC includes a wide variety of benchmark kernels used to measure accelerator performance in both single node and cluster configurations. The hybrid model implementation attempts to place all major computation on accelerator devices, and uses MPI to synchronize and aggregate results. In some cases, MPI Groups are used to gradually reduce the number of accelerators used for computation as the problem size drops. Porting this behavior to OpenSHMEM required implementing several synchronizing collective operations, and using SHMEM teams to replace MPI Group functionality. Benchmark results on a Cray XK7 system with one GPU per compute node show that SHMEM performance is equal to MPI performance in these hybrid tasks. These results and porting experience show that using OpenSHMEM for accelerator devices benefits from adding functionality for synchronization and teams, and would further benefit from adding support for communication within accelerator kernels. (Notice: This manuscript has been authored by UT-Battelle, LLC under Contract No. DE- AC05-00OR22725 with the U.S. Department of Energy. The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan). This research used resources of the Center for Computational Sciences at Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. De-AC05-00OR22725.)

Synthetic graph generation for data-intensive HPC benchmarking: Scalability, analysis and real-world application

The benchmarking effort within the Extreme Scale Systems Center at Oak Ridge National Laboratory seeks to provide High Performance Computing benchmarks and test suites of interest to the DoD sponsor. The work described in this report is a part of the effort focusing on graph generation. A previously developed benchmark, SystemBurn, allows the emulation of a broad spectrum of application behavior profiles within a single framework. To complement this effort, similar capabilities are desired for graph-centric problems. This report described the in-depth analysis of the generated synthetic graphs’ properties at a variety of scales using different generator implementations and examines their applicability to replicating real world datasets.

Graph Generator Survey

The benchmarking effort within the Extreme Scale Systems Center at Oak Ridge National Laboratory seeks to provide High Performance Computing benchmarks and test suites of interest to the DoD sponsor. The work described in this report is a part of the effort focusing on graph generation. A previously developed benchmark, SystemBurn, allowed the emulation of dierent application behavior profiles within a single framework. To complement this effort, similar capabilities are desired for graph-centric problems. This report examines existing synthetic graph generator implementations in preparation for further study on the properties of their generated synthetic graphs.