Architecture aware parallel programming in Glasgow parallel Haskell (GPH)

Abstract

General purpose computing architectures are evolving quickly to become manycore and hierarchical: i.e. a core can communicate more quickly locally than globally. To be effective on such architectures, programming models must be aware of the communications hierarchy. This thesis investigates a programming model that aims to share the responsibility of task placement, load balance, thread creation, and synchronisation between the application developer and the runtime system. The main contribution of this thesis is the development of four new architectureaware constructs for Glasgow parallel Haskell that exploit information about task size and aim to reduce communication for small tasks, preserve data locality, or to distribute large units of work. We define a semantics for the constructs that specifies the sets of PEs that each construct identifies, and we check four properties of the semantics using QuickCheck. We report a preliminary investigation of architecture aware programming models that abstract over the new constructs. In particular, we propose architecture aware evaluation strategies and skeletons. We investigate three common paradigms, such as data parallelism, divide-and-conquer and nested parallelism, on hierarchical architectures with up to 224 cores. The results show that the architecture-aware programming model consistently delivers better speedup and scalability than existing constructs, together with a dramatic reduction in the execution time variability. We present a comparison of functional multicore technologies and it reports some of the first ever multicore results for the Feedback Directed Implicit Parallelism (FDIP) and the semi-explicit parallelism (GpH and Eden) languages. The comparison reflects the growing maturity of the field by systematically evaluating four parallel Haskell implementations on a common multicore architecture. The comparison contrasts the programming effort each language requires with the parallel performance delivered. We investigate the minimum thread granularity required to achieve satisfactory performance for three implementations parallel functional language on a multicore platform. The results show that GHC-GUM requires a larger thread granularity than Eden and GHC-SMP. The thread granularity rises as the number of cores rises.