Abstract:
Many modern systems provide heterogeneous parallelism, for example NUMA multi-core processors and CPU-GPU combinations. Placement, scheduling and indeed algorithm choices affect the overall execution time and, for portable programs, must adapt to the target machine at either load-time or run-time. We see these choices as preserving I/O determinism but exposing performance non-determinism. We use Petri-nets as an intermediate representation for programs to give a unified view of all forms of performance non-determinism. This includes some scenarios which other models cannot support. Whilst NP-hard, efficient heuristics for approximating optimum executions in these nets would lead to performant portable execution across arbitrary heterogeneous architectures.

Paper available at ACM.