Final Program

For decades parallel computing has been the focus of intense research and development in selected fields, and numerous large-scale parallel applications have been developed. SPMD via MPI has been a dominant approach to parallelism to date, but this approach alone will be insufficient going forward. Presently we are on the threshold of mass deployment of parallelism across most application areas, but the path to developing these applications is uncertain. There are many programming models, languages and architectures from which to pick, and the number of choices is growing. In this presentation we discuss some of the principles of parallel application development that have produced today's codes, and how we can address these principles going forward. We also discuss what must change in order to move forward and give ideas for developing parallel applications now that will have sustained value in the future.

[ slides ]

The OpenMP API provides a portable model for efficient, high level thread-parallel programming across platforms, vendors, operating systems. We want a model with the same advantages to address compute accelerators. In this talk, we explore today's accelerator landscape, along with the perils of current programming methods. We demonstrate why OpenCL, while impressive and important, doesn't already solve the programming problem. We close with a summary of the PGI Accelerator Model, and discuss some of the implementation issues into which we have run.

Although multithreaded programming languages are common, there has been a surprising amount of confusion surrounding the basic meaning of shared variables. This has spanned various programming languages and shared memory parallel programming APIs. There is finally a growing consensus, both that programming languages should by default guarantee an interleaving-based semantics (sequential consistency) for data-race-free programs, and on what that should mean. We discuss the motivation for, and consequences of, this guarantee. We briefly survey where various programming languages stand with respect to this guarantee.

The memory model description used in OpenMP 3.0 is largely compatible with such an approach, in spite of the very different terminology used in the specification. We argue however that the current reliance on explicit and implicit "flush" operations has some unfortunate consequences. In particular, the current atomics directives are both relatively difficult to use and their mere existence may have unexpected performance consequences, even if they're not used at all. We propose that future OpenMP specifications follow a somewhat different model, more similar to that used in the current draft C and C++ standards.

[ slides ]