Automatic Scheduling for Cache Only Memory Architectures

For parallel and distributed systems to gain wider acceptance than they have to date, they must become significantly easier to program. Fundamentally, parallel programming is more difficult than sequential programming as long as data and computation must be distributed by the programmer. Cache Only Memory Architectures (COMAs) provide a Distributed Shared Memory (DSM) where data distribution is performed automatically and transparently. This paper generalizes this idea to achieve the same distribution for computation, thus arriving at an automatic and transparent form of scheduling. Once computation distribution (scheduling) is as automatic and transparent as data distribution, parallel computers become approximately as easy to program as sequential computers. Where COMA literature normally makes no assumptions concerning the parallel programs which use the DSM, we use special compiler techniques originally developed for multithreaded and dataflow architectures. Having done so, we can specify ways of significantly simplifying the basic COMA coherency protocols, while at the same time enabling automatic, transparent, adaptive run-time scheduling. This work was supported in part with funds of the Deutsche Forschungsgemeinschaft under reference number WA 357/11-2 within the priority program “System and Circuit Technology for Massively Parallel Computation”.