Browsing by Author "Hollingsworth, Jeffrey K."

Now showing 1 - 8 of 8
  • Thumbnail Image
    Item
    Active Harmony: Towards Automated Performance Tuning
    (2002-08-01) Tapus, Cristian; Chung, I-Hsin; Hollingsworth, Jeffrey K.
    In this paper we present the Active Harmony automated runtime tuning system. We describe the interface used by programs to make applications tunable. We present the Library Specification Layer which helps program library developers expose multiple variations of the same API using different algorithms. The Library Specification Language helps to select the most appropriate program library to tune the overall performance. We also present the optimization algorithm that we used to adjust parameters in the application and the libraries. Finally, we present results that show how the system is able to tune several real applications. The automated tuning system is able to tune the application parameters to within a few percent of the best value after evaluating only 11 configurations out of over 1,700 possible combinations. Also UMIACS-TR-2002-54
  • Thumbnail Image
    Item
    Automated Cluster-Based Web Service Performance Tuning
    (2003-09-02) Chung, I-Hsin; Hollingsworth, Jeffrey K.
    In this paper, we apply the Active Harmony system to improve the performance of a cluster-based web service system. The performance improvement cannot easily be achieved by tuning individual components for such a system. The experimental results show that there is no single configuration for the system that performs well for all kinds of workloads. By tuning the parameters, the Active Harmony helps the system adapt to different workloads and improve the performance up to 16%. For scalability, we demonstrate how to reduce the time when tuning a large system with many tunable parameters. Finally an algorithm is proposed to automatically adjust the structure of cluster-based web systems, and the system throughput is improved up to 70% using this technology. (UMIACS-TR-2003-84)
  • Thumbnail Image
    Item
    Grindstone: A Test Suite for Parallel Performance Tools
    (1998-10-15) Hollingsworth, Jeffrey K.; Steele, Michael
    We describe Grindstone, a suite of programs for testing and calibrating parallel performance measurement tools. The suite consists of nine simple SPMD style PVM programs that demonstrate common communication and computational bottlenecks that occur in parallel programs. In addition, we provide a short case study that demonstrates the use of the test suite on three performance tools for PVM. The results of the case study showed that we were able to uncover bugs or other anomalies in all three tools. The paper also describes how to acquire, compile, and use the test suite. (Also cross-referenced as UMIACS-TR-96-73)
  • Thumbnail Image
    Item
    Load Balancing Factor (LBF): A Workload Migration Metric
    (2003-08-01) Eom, Hyeonsang; Hollingsworth, Jeffrey K.
    We introduce a new performance metric, called Load Balancing Factor (LBF), to evaluate different tuning alternatives of workload migration within a distributed/parallel program. The metric is unique because it shows the performance implications of a specific tuning alternative rather than quantifying where time is spent in the program. Previously we developed a variation of the metric for coarse-grained process placement, and demonstrated that it accurately predicts the placement impact. In this paper we focus on a variation designed for fine-grained function shipping in a client/server environment and present its online algorithm. We use a synthetic application to show that LBF provides accurate guidance about procedure-level migration. (UMIACS-TR-2002-55)
  • Thumbnail Image
    Item
    A Pilot Study to Evaluate Development Effort for High Performance Computing
    (2004-04) Basili, Victor R.; Asgari, Sima; Hochstein, Lorin; Hollingsworth, Jeffrey K.; Shull, Forrest; Zelkowitz, Marvin V.
    The ability to write programs that execute efficiently on modern parallel computers has not been fully studied. In a DARPA-sponsored project, we are looking at measuring the development time for programs written for high performance computers (HPC). To attack this relatively novel measurement problem, our goal is to initially measure such development time in student programming to evaluate our own experimental protocols. Based on these results, we will generate a set of feasible experimental methods that can then be applied with more confidence to professional expert programmers. This paper describes a first pilot study addressing those goals. We ran an observational study with 15 students in a graduate level High Performance Computing class at the University of Maryland. We collected data concerning development effort, developer activities and chronology, and resulting code performance, for two programming assignments using different HPC development approaches. While we did not find strong correlations between the expected factors, the primary outputs of this study are a set of experimental lessons learned and 12 wellformed hypotheses that will guard future study.
  • Thumbnail Image
    Item
    Rate Windows for Efficient Network and I/O Throttling
    (2000-07-12) Ryu, Kyung D.; Hollingsworth, Jeffrey K.; Keleher, Peter J.
    This paper proposes and evaluates a new mechanism for I/O and network rate policing. The goal of the proposed system is to provide an simple, yet effective way to enforce resource limits on target classes of jobs in a system. The basic approach is useful for several types of systems including running background jobs on idle workstations, and providing resource limits on network intensive applications such as virtual web server hosting. Our approach is quite simple, we use a sliding window average of recent events to compute the average rate for a target resource. The assigned limit is enforced by forcing application processes to sleep when they issue requests that would bring their resource utilization out of the allowable profile. Our experimental results that show that we are able to provide the target resource limitations within a few percent, and do so with no measurable slowdown of the overall system. (Also cross-referenced as UMIACS-TR-2000-53)
  • Thumbnail Image
    Item
    Using Content-Derived Names for Caching and Software Distribution
    (1998-10-15) Hollingsworth, Jeffrey K.; Miller, Ethan L.
    Maintaining replicated data in wide area information services such as the World Wide Web is a difficult problem. Ensuring that the correct versions of libraries and images are installed for application programs presents similar challenges. In this paper, we present a simple scheme to facilitate both of these tasks using content-derived names (CDNs). Content-based naming uses digital signatures to compute a name for an object based only on its content. CDNs can be applied to several common problems of modern computer systems. Caching on the World Wide Web is simplified by allowing references to an object by its content rather than just its location. In a similar fashion, applications can request library objects by their content without having to rely on the presence of a file system hierarchy that the application recognizes. Further, applications that require different versions of an object can coexist peacefully on the same machine. While this idea is still in its early stages, we present experimental evidence from a study of World Wide Web objects that indicates that CDNs could reduce network traffic by allowing requests to be satisfied by differently-named duplicates with the same contents. (Also cross-referenced as UMIACS-TR-96-55)
  • Thumbnail Image
    Item
    ViPEr-HiSS: A Case for Storage Design Tools
    (1999-10-27) Golubchik, Leana; Dunnick, Joseph; Hollingsworth, Jeffrey K.
    The viability of large-scale multimedia applications, depends on the performance of storage systems. Providing cost-effective access to vast amounts of video, image, audio, and text data, requires (a) proper configuration of storage hierarchies as well as (b) efficient resource management techniques at all levels of the storage hierarchy. The resulting complexities of the hardware/software co-design in turn contribute to difficulties in making accurate predictions about performance, scalability, and cost-effectiveness of a storage system. Moreover, poor decisions at design time can be costly and problematic to correct in later stages of development. Hence, measurement of systems after they have been developed is not a desirable approach to predicting their performance. What is needed is the ability to evaluate the system's design while there are still opportunities to make corrections to fundamental design flaws. In this paper we describe the framework of ViPEr-HiSS, a tool which facilitates design, development, and subsequent performance evaluation of designs of multimedia storage hierarchies by providing mechanisms for relatively easy experimentation with (a) system configurations as well as (b) application- and media-aware resource management techniques. (Also cross-referenced as UMIACS-TR-99-69)