Wednesday

Half Day of Hands on Demonstrations

Agenda for Wednesday(3/22/06):

Continental Breakfast: 4185 A.V. Williams
8:00am - 9:00am

Demos: 9:00am - Noon.

Schedule
Updated March 20, 2006

• Active Harmony
  
  Ananta Tiwari, University of Maryland
  Room 4140

  Active Harmony framework permits automated runtime adaptation of
  algorithms, data distribution and load balancing for parallel
  applications. The framework exploits the tuning options specified
  by parallel applications and adapts computation online based on
  observed performance and changing conditions. In this demo, we will
  present the Active Harmony system and describe the interface
  used by programs to make applications tunable. We will use real
  scientific applications for this purpose.

• Dyninst on Threaded
  Applications Summary
  Matt Legendre, University of Wisconsin
  Room 4161

  We will demonstrate Dyninst's new capacity to work with threaded
  applications, such as Microsoft Word, by using it to build a system
  call tracer.

• Kernel Profiles for
  User Applications
  Via KerninstAPI
  Mike Brim, University of Wisconsin
  
  Room 4140

  This demonstration will highlight the ability of a KerninstAPI-based
  profiler to dynamically observe the functions in the Linux kernel
  executed on behalf of a given user application, and subsequently
  produce a call graph with execution counts. Profiling via instrumentation
  can often capture behavior not visible when observing execution using
  periodic program counter sampling, as is used by tools such as OProfile.

• Root Cause Analysis of
  Failures In Large-Scale
  Computing Environments
  Alex Mirgorodskiy, University of Wisconsin

  Room 4140

  We demonstrate how self-propelled instrumentation can be combined
  with machine-learning techniques to locate anomalies in distributed
  environments. We use self-propelled instrumentation to collect
  function-level traces for all nodes of a parallel application. When one
  of the application nodes crashes or freezes, we recover collected traces
  and save them to disk. Next, we identify the failed node by comparing
  traces to each other with an automated outlier-detection technique.
  Finally, we identify a function that is a likely cause of the failure
  and report it to the analyst.

• Multiplatform Paradyn
  Todd Miller, University of Wisconsin
  Room 4140

  We demonstrate Paradyn instrumenting an application with processes
  running on IA-64, x86, and x86-64 machines. We let the Performance
  Consultant automatically determine the nature of the bottleneck and
  use the visualization tools to demonstrate its periodicity.

• Grid Resource Discovery
  Using MRNet and the
  Network Weather Service
  Dorian Arnold, University of Wisconsin
  Room 4140

  MRNet is a tree-based overlay network for scalable, efficient data
  multicast and reduction services. The Network Weather Service (NWS)
  monitors and dynamically forecasts the performance of network and
  computational resources. This demonstration showcases the integration
  of MRNet and the Network Weather Service into a matchmaking service
  that could be used for applications like Grid scheduling. The MRNet
  front-end allows users to specify resource constraints(e.g. Available
  CPU > X MFLOPS && Available Memory > Y MB). NWS forecasts are computed
  at the MRNet leaves which stream the results up the MRNet tree. Filters
  in the MRNet tree execute constraint solvers that propagate the names
  and properties of the resources that match the queries constraints.

• ParseThat - Exercising
  the Dyninst Library
  Ray Chen, University of Maryland
  Room 4139

  We will demonstrate ParseThat, a tool that exercises the Dyninst
  library on arbitrary binaries. A variety of targets will be
  instrumented, including single-threaded, multi-threaded, and stripped
  binaries. Target binaries will be accepted on request.