Haskell on a shared-memory multiprocessor

Multi-core processors are coming, and we need ways to program them. The combination of purely-functional programming and explicit, monadic threads, communicating using transactional memory, looks like a particularly promising way to do so. This paper describes a full-scale implementation of shared-memory parallel Haskell, based on the Glasgow Haskell Compiler. Our main technical contribution is a lock-free mechanism for evaluating shared thunks that eliminates the major performance bottleneck in parallel evaluation of a lazy language. Our results are preliminary but promising: we can demonstrate wall-clock speedups of a serious application (GHC itself), even with only two processors, compared to the same application compiled for a uni-processor. Categories and Subject Descriptors D.3.2 [Language Classifications]: Applicative (functional) languages General Terms Languages, Performance

• Type: Conference
• Year: 2005
• Where: HASKELL
• Authors: Tim Harris, Simon Marlow, Simon L. Peyton Jones
• Keywords: Glasgow Haskell Compiler, HASKELL 2005, Major Performance Bottleneck, Shared-memory Parallel Haskell