The Chaos project has developed methods to map a broad range of
challenging applications onto high performance computer architectures.
A major focus of this work is development of parallelization
techniques for irregular scientific problems -- problems that are
unstructured, sparse, adaptive or block-structured. This project
works extensively with application
developers in many disciplines and with parallel compiler vendors.
Many concepts first described and prototyped by this
project are making their way into High Performance
Fortran during the ongoing second round of language definition.
This work is also leading to the development of runtime support (Meta-Chaos
) to couple runtime libraries used in data- and task-parallel
compilers. Meta-Chaos is a central component for the common compiler
runtime support being developed by the Parallel Compiler Runtime
Consortium.
We are currently developing techniques that will allow parallel compute and data objects to offer their services to remotely connected clients. The goal is to develop techniques that will make it possible to compose programs running on any combination of distributed memory, shared memory or networked microcomputers or workstations. We are motivating this research by software interoperability scenarios associated with two classes of applications. The first class is sensor data processing and integration, and the second is complex physical simulations. We have developed early prototypes of our data parallel program coupling software and have employed our prototype to demonstrate the ability to couple separately executing High Performance Fortran programs, and to couple High Performance Fortran programs with applications developed using the Maryland CHAOS and Multiblock Parti libraries.
Based upon our experiences in developing runtime libraries and parallelizing applications, we have developed several compilation techniques. Our goal is to be able to automate our hand parallelization/optimization techniques through the use of compilers. We have employed the Fortran~D compilation system (developed primarily at Rice University) as the infrastructure for implementation of our techniques.
For generating efficient code from applications having multiple levels of indirection, we have developed an index flattening technique (based upon the notion of program slicing). This technique transforms a loop having multiple levels of indirection into a series of loops having at most a single level of indirection. We have also observed that aggressive interprocedural optimizations are required to deal with large applications that have irregular accesses to data or large I/O requirements. We have developed an Interprocedural Partial Redundancy Elimination (IPRE) technique for performing interprocedural placement of communication preprocessing and collective communication statements. We are currently working on Interprocedural Balanced Code Placement (IBCP), which will allow us to overlap computation and communication across procedure boundaries. We are also working on generating distributed memory code from Fortran~90 codes that use pointers and recursive data structures. -by Joel Saltz