Software Infrastructure for Application Development

1
Software Infrastructurefor Application
Development

• Laxmikant Kale
• David Padua
• Computer Science Department

2
Computer Science Roles

• Work closely with mathematicians and engineers
• Give engineers state-of-the-art tools
• Provide benchmark for Java development
• Symbolic support of FE software generation
• Interaction with mathematicians and engineers
• HPC environment
• Java compiler
• Runtime infrastructure

3
HPC Environment
Symbolic algebra support
Fortran code
Container Semantics
Java Compiler
F2J
Parallel Infrastructure
Uniprocessor
4
Why Java?

• Object model
• Modular and reusable code development
• Amortize the high cost of software development
  for high performance scientific codes
• Much faster to debug and develop
• Disciplined memory model
• garbage collection
• Java Grande
• Industry trend
• May replace C

5
Runtime environments

• Object-based runtime infrastructure
• Provide parallel support libraries
• past work parallel Java, object-based load
  balancing
• We have developed parallel Java
• Object based parallelism (not just threads)
• Fast asynchronous remote method invocation
• Substantially faster than RMI
• Can inter-operate with other languages/paradigms
• Implemented in Converse framework
• modules written in Java, MPI-C, Charm etc. can
  co-exist
• PDPTA -97 Design and implementation of
  parallel Java

6
Object-based load balancing

• Charm supports object based load balancing
• Parallel C
• Dynamic and periodic balancing
• seed movement
• migration mechanisms with automatic forwarding
• migration strategies
• Used in molecular dynamics application
• Used for cluster adaptivity
• Can be adapted for Java

7
Load balancing before
8
Load balancing after
9
Java Compiler

• Oriented to the needs of the project
• Targets uniprocessors and parallel machines
• Use advanced analysis techniques
• Builds on expertise in compilers
• Promise of Java
• Extensive static analysis

10
Container Class analysis

• Goal
• To optimize symbolic Java applications with
  extensive use of container classes or numerical
  Java codes using sparse matrices
• Program optimization
• Hard-wire semantics of Java container classes
  into the compiler
• java.util.Vector
• java.util.Hashtable
• sparse matrix

11
Container analysis contd.

• Design optimization techniques to exploit their
  semantic properties
• access analysis
• commutativity analysis
• exploit associativity
• loop parallelization
• pointer analysis

12
Performance
CPSD
13
Performance
CPSD
14
Ongoing and future work

• Develop Java compiler for scientific applications
• Fortran to Java translator
• Interactions with IBM researchers
• Interested in compilation issues
• Grad student exchanges
• Consultation on engineering applications
• Optimization and parallelization
• Code organization
• Work with mathematicians
• Symbolic analysis
• Compiler strategies