Dynamic Chunking in Eden

Parallel programming generally requires awareness of the granularity and communication requirements of parallel subtasks, since without precaution, the overhead for parameter and result communication may outweigh the gain of parallel processing. While this problem is often solved explicitly at the language level, it can also be alleviated by optimising message passing mechanisms in the runtime environment. We describe how a simple buffering mechanism introduces dynamic list chunking in the runtime environment of the parallel functional language Eden. We discuss design and implementation aspects of dynamic chunking and compare its effects to the original version in a set of measurements. Our optimisation is justified by a simple cost model, measurements analyse the overhead and illustrate the impact of the changed message passing mechanism.