high performance computing on graphics processing units: hgpu.org

Future embedded systems for performance-demanding applications will be massively parallel. High performance tasks will be parallel programs, running on several cores, rather than single threads running on single cores. For hard real-time applications, WCETs for such tasks must be bounded. Low-level parallel programming models, based on concurrent threads, are notoriously hard to use due to their inherent nondeterminism. Therefore the parallel processing community has long considered high-level parallel programming models, which restrict the low-level models to regain determinism. In this position paper we argue that such parallel programming models are beneficial also for WCET analysis of parallel programs. We review some proposed models, and discuss their influence on timing predictability. In particular we identify data parallel programming as a suitable paradigm as it is deterministic and allows current methods for WCET analysis to be extended to parallel code. GPUs are increasingly used for high performance applications: we discuss a current GPU architecture, and we argue that it offers a parallel platform for compute-intensive applications for which it seems possible to construct precise timing models. Thus, a promising route for the future is to develop WCET analyses for data-parallel software running on GPUs.