Performance Evaluation of Scalable Parallel Systems

Proposed Activities

Architecture, Algorithm & Workload Characterization

• Synthetic benchmarks of architectural features and low level parallel characteristics
• High-performance scalable algorithms (e.g. FFTs, computational linear algebra)
• Strategic applications

Instrumentation

Need to Gather Data at Many Levels

Performance Prediction

• Compile Time

The Parts

• CTC : comprehensive multi-dimensional workload characterization; limited predictive abilities

The Parts: UMD

• Dynamic Instrumentation of Runtime Processes
  • modify binary of executing program
  • only incur overhead for desired instrumentation
  • requires operational systems and applications
• Example
  • report messages sent by foo (or its children)

The Parts: UCLA

• parallel simulation of parallel applications using coarse level component models

The Parts: NMSU

• detailed memory hierarchy models & uniprocessor multi-task workloads

The Whole

Alliance infrastructure will support detailed performance characterization of strategic applications/algorithms on future parallel systems:

• ability to instrument parallel system at multiple levels
• interface instrumentation framework w/ simulator
• sophisticated multi-level simulation capability
• integration with capacity planning via ARMS interface

Impact on Infrastructure (1)

• Application developers
  • improve efficiency of (strategic) applications on existing architectures
• Parallel system (software & hardware) designers
  • define next generation of algorithms & architectures

Impact on Infrastructure (2)

• System operations
  • guide globally efficient resource allocation via resource management systems like ARMS
  • optimum application to architecture mapping in heterogeneous environments
• Supercomputing community
  • data & tool repository will be an invaluable asset for the community

Deliverables

• Primary Deliverable: The interactive, web-accessible data & model repository
  • hierarchy of strategic applications, algorithms, & synthetic benchmarks
  • application scalability studies
  • performance studies of alternative designs for architectural or system software components:
    • switch protocols & architectures
    • message passing libraries
  • performance daemon for ARMS
  • etc.

Updated: October 30, 1996