The Case for Binary Rewriting at Runtime for Efficient Implementation of High-Level Programming Models in HPC

摘要

Implementations of Parallel Programming Models are provided either as language extensions, completely new languages or as a library. The first two options often provides high productivity, but requires the porting of codes. In contrast, calls to new libraries can be added more easily, however the use of abstractions in such programming model implementations can have high runtime overhead. In both cases, the mentioned drawbacks often hinder the adaptation of novel programming models for large existing codes. To combine the advantages of compiler analysis with the composability of pure libraries towards more efficient programming model implementations, in this paper, we propose a low level API for programmer controlled binary rewriting at runtime. This can be used by programming models provided as libraries to efficiently integrate their abstractions with application code. It enables incremental adoption for existing codes as well as favoring input-dependent optimization strategies yet providing similar performance as language extension approaches. We show first promising experiences.