This paper explores the idea of the processor as an asynchronous network, called the micronet, of functional units which compute concurrently and communicate asynchronously. A micronet-based asynchronous processor exposes spatial as well as temporal concurrency. We analyse the performance of the 'processor-as-a-network' by comparing three scheduling algorithms for exploiting Instruction Level Parallelism (ILP). Schedulers for synchronous architectures have relied on deterministic instruction execution times. In contrast, ILP scheduling in micronet-based architectures is a challenge as it is less certain in advance when instructions start execution and when results become available. Performance results comparing the three schedulers are presented for SPEC95 benchmarks executing on a cycle-accurate model of the micronet architecture.
展开▼