Adapt or Become Extinct! The Case for a Unified Framework for Deployment-Time Optimization

Georgios Goumas, Sally A. McKee, Magnus Själander, Thomas R. Gross, Sven Karlsson, Christian W. Probst, Lixin Zhang

    Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

    307 Downloads (Pure)

    Abstract

    The High-Performance Computing ecosystem consists of a large variety of execution platforms that demonstrate a wide diversity in hardware characteristics such as CPU architecture, memory organization, interconnection network, accelerators, etc. This environment also presents a number of hard boundaries (walls) for applications which limit software development (parallel programming wall), performance (memory wall, communication wall) and viability (power wall). The only way to survive in such a demanding environment is by adaptation. In this paper we discuss how dynamic information collected during the execution of an application can be utilized to adapt the execution context and may lead to performance gains beyond those provided by static information and compile-time adaptation. We consider specialization based on dynamic information like user input, architectural characteristics such as the memory hierarchy organization, and the execution prole of the application as obtained from the execution platform's performance monitoring units. One of the challenges of future execution platforms is to allow the seamless integration of these various kinds of information with information obtained from static analysis (either during ahead-of-time or just-in-time) compilation. We extend the notion of information-driven adaptation and outline the architecture of an infrastructure designed to enable information ow and adaptation throughout the life-cycle of an application.
    Original languageEnglish
    Title of host publicationEXADAPT '11 Proceedings of the 1st International Workshop on Adaptive Self-Tuning Computing Systems for the Exaflop Era
    PublisherUniversity of Strathclyde
    Publication date2011
    ISBN (Print)978-1-4503-0708-6
    DOIs
    Publication statusPublished - 2011
    EventInternational Workshop on Adaptive Self-Tuning Computing Systems for the Exaflop Era - San Jose, California
    Duration: 1 Jan 2011 → …
    Conference number: 1

    Conference

    ConferenceInternational Workshop on Adaptive Self-Tuning Computing Systems for the Exaflop Era
    Number1
    CitySan Jose, California
    Period01/01/2011 → …

    Bibliographical note

    © ACM, 2011. This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in EXADAPT '11, http://doi.acm.org/10.1145/2000417.2000422

    Fingerprint Dive into the research topics of 'Adapt or Become Extinct! The Case for a Unified Framework for Deployment-Time Optimization'. Together they form a unique fingerprint.

    Cite this