Profiling Facility on a Metasystem

1
Profiling Facility on a Metasystem

• Carlos Figueira and Emilio Hernández
• Universidad Simón Bolívar
• Caracas, Venezuela

2
Metasystems

• Heterogeneous computational resources tied by
  middleware, and accessible transparently through
  a network
• Execution nodes can be chosen by a scheduler
  according to application requirements, and
  platform availability and capabilities

3
Motivation

• How can we help users understand their
  applications performance?
• Providing relevant information to tune their
  applications
• Yet what does tuning an application for a
  metacomputer actually mean?

4
Profiling on a metasystem

• An application profile gives basic information
  about execution performance discriminated by
  components
• Two factors compromise usefulness of classic
  profiling on metasystems
• Heterogeneity
• External perturbations

5
Profiling Levels

• Condensed profiles attempt to provide basic
  information for platform independent tuning. Its
  normalized, according to format and relative
  computing power, and lightweight.
• Complete profiles depend on available tools and
  are platform dependent

6
SUMA

• Metasystem aimed at executing high performance
  Java byte code applications
• Implemented on CORBA, includes clients (running
  on user machines), execution agents (controling
  execution at metasystem nodes) and core
  components (e.g., scheduler)

7
Profiling in SUMA
8
Current prototype

• Back-end profiler HPROF
• Condensed profile object
• Performance information of most relevant methods,
  including self and inclusive times
• Execution node data (normalized power,
  architecture, OS and JVM versions, etc)
• Deployed campus-wide

9
Experiments

• Sequential application (Jess) allocated to dual
  processor Pentium III, 600 MHz with Linux (kernel
  2.2.14smp), JDK 1.2.2
• Reported correct order of most important methods

10
Conclusions

• General, adaptable model for building execution
  profiles in a Java-based metasystem
• Useful for both application and metasystem
  performance analysis
• Future work extension to parallel applications