Compiling Assembly Pattern on a Shared Virtual Memory

Abstract

Among the variety of tools in a programming environment (debugger, performance analyzer, data mapping assistant, user interface, compiler, runtime support), we focus here on the compiler item. The definition of machine-independent parallel programming languages is a relevant example of the need for portability. However, we believe that rather than focusing exclusively on data distribution approaches, we may merge both data distribution and control distribution to target a wider range of applications. To rely on control distribution, we need a global address space, which, on scalable multiprocessors, may be afforded by a Shared Virtual Memory (SVM). Thanks to the SVM, the code generation process is much simpler: there is no need to an explicit data distribution, nor localization of array subscripts, nor management of non-local data buffers, nor communication schedules generation (with message vectorization, coalescing, aggregation). All this is handled automatically by the runtime support. Furthermore, the compilation of procedure and function calls within a parallel loop becomes easily feasible. Incremental (step by step) parallelization of codes thanks to the global address space is an interesting advantage. The drawback is the coherence grain which is not always well suited to the application.