Workshop Agenda :: Workshop slides

Organizers

• Barton Miller (University of Wisconsin - Madison) "bart" AT "cs.wisc.edu"
• John Mellor-Crummey (Rice University) "johnmc" AT "cs.rice.edu"

Abstract

Soon petascale computer systems will be available to the DOE science community. Compared to existing systems, these emerging machines represent a dramatic increase in scale and complexity. Making effective use of systems at this scale will require performance tools that help application developers by pinpointing and explaining application performance bottlenecks. Two performance challenges for applications are to make effective use of processor nodes, and to keep parallelization overhead low enough so that extreme-scale parallelism is practical.

Providing tools that make it possible to address these problems is requiring increasingly sophisticated methods for instrumentation, measurement, analysis and modeling of application performance. Meeting these challenges will require tools with unprecedented capabilities.

Today, research and development in the area of performance tools for parallel systems is fragmented. Individual groups have developed tools and components with unique capabilities. Working together to capitalize these capabilities into shared community infrastructure would reduce duplication of effort, accelerate development of effective tools with increasing sophistication, and simplify the process of developing tools for new systems as they emerge.

The goal of this workshop was to bring together performance tools researchers to foster collaboration on a community infrastructure for performance tools. Specific aims of the workshop included

• Identifying key sets of functionality for performance tools.
• Surveying available research and commercial tools in this space.
• Proposing a component taxonomy/architecture for tools
• Identifying common needs, functionality, and opportunities for sharing infrastructure.
• Identifying areas for standardization of interfaces and functionality
• Working as a community to draft preliminary standards to facilitate component reuse.

Agenda

Day 1 - Monday, July 16

• Welcome, Introductions, Workshop Challenges and Goals
  John Mellor-Crummey, Rice University
  Bart Miller, University of Wisconsin
• The Deconstruction of Dyninst
  Bart Miller, University of Wisconsin
  Symtab API (pdf coming soon), Stackwalker API (pdf)
• The Linux Perfmon Interface: Overview and Challenges
  Stefane Eranian, HP
• Analyzing Memory Hierarchy Performance Using AMD Opteron Performance Counters
  John McCalpin, AMD
• Sampling-Based Measurement and Analysis
  John Mellor-Crummey, Rice University
• Open|SpeedShop Capabilities and Internal Structure: Current/Petascale
  Jim Galarowicz, Krell Institute
• Scalable Performance Analysis of Large-Scale Applications
  Felix Wolf, Forschungszentrum Julich
• Scalability of Trace Analysis Tools
  Jesus Labarta, European Center for Parallelism of Barcelona

Day 2 - Tuesday, July 17

• What is Scalable Instrumentation?
  Robert Cohn, Intel
• Performance Evaluation Using the TAU Performance System
  Sameer Shende, University of Oregon
• Metadata Collection for Performance Analysis
  Karen Karavanic, Portland State University
• General discussion to identify candidates for groups
• "Hot Button" Issues Short Presentations

Day 3 - Wednesday, July 18

• Working groups meet for selected interface proposals.

Day 4 - Thursday, July 19

• Morning Working groups meet for selected interface proposals.
• Afternoon Presentation of proposals and feedback. Plan for ongoing work.

Sponsors