Maestro: a self-organizing peer-to-peer dataflow framework using reinforcement learning
Proceedings of the 18th ACM international symposium on High performance distributed computing
Hi-index | 0.00 |
Dataflow computation models enable simpler and more efficient management of the memory hierarchy - a key barrier to the performance of many parallel programs. This paper describes a dataflow language based on Java. Use of the dataflow model enables a programmer to generate parallel programs without explicit directions for message passing, work allocation and synchronization. A small handful of additional syntactic constructs are required. A pre-processor is used to convert Dataflow Java programs to standard portable Java.The underlying run-time system was easy to implement using Java's object modeling and communications primitives. Although raw performance lags behind an equivalent C-based system, we were able to demonstrate useful speedups in a heterogeneous environment, thus amply illustrating the potential power of the Dataflow Java approach to use all machines - of whatever type - that might be available on a network .. when Java JIT compiler technology matures.