Snir warned against creating a hybrid language based on the work of the three companies. "I don't think the first version of a language can be designed by committee. It has to have its own personality," he said.
A more evolutionary alternative, he said, would be to move to emerging languages such as Co-Array Fortran or Universal Parallel C. Those languages require smaller advances in compiler technology, Snir said, but they have gotten little traction thus far.
Neither Darpa nor the three companies are providing many details about the competing languages. Snir said Cray's Chapel is the closest to MPI because it is focused on data parallelism. IBM's X10 introduces new ideas such as atomic data structures and asynchronous processes, he said. Sun's Fortress, which has been described as "Matlab on steroids," would require the most aggressive changes in compilers of the trio, Snir said.
Sun's Mitchell said Fortress would be open and independent of any operating system. It uses parallel programming as its default mode and can incorporate new libraries as if they were native to the language. That feature is prized in high-performance computing, where developers routinely employ third-party software libraries.
Sun is also exploring compilers for Fortress that would use both static and dynamic run-time optimizations to increase parallelism, Mitchell said.
Cray's Chapel would be made available to the industry if Darpa selects it. "Some of the work we do in the system software area will be contributed back to the open-source community, but I am not aware of any major pieces of software at this time that we are planning to release," said Steve Scott, chief technology officer at Cray.
Cray is also working with third-party tool developers, which would be free to make widely available any programs that develop around Chapel, he added.
A spokesman for IBM would not make any details of X10 available. He said that the company had not decided whether it would make the language available for other systems makers to use.
Snir said the problems finding a replacement for MPI are "less technical and more commercial. We have not figured out a business model to cover the expenses of building such an advanced development platform."
MPI has demonstrated the potential to scale up to large systems; indeed, it is already deployed in a version of IBM's BlueGene/L supercomputer that uses 128,000 processors, Snir said. However, such large systems do create great programming complexity in MPI.
The MPI libraries, developed in the early 1990s and now in a version 2.0 release, effectively have no road map.
"We may reconvene to write minor enhancements, but we don't expect MPI to evolve significantly," Snir remarked. "MPI's lifetime has already exceeded my expectations."
