Hitchhiker ’ s Guide to GUM

This report documents the process of extending the GUM run-time system (RTS) for Glasgow parallel Haskell (GpH). First, we obtain the sources, set up the environment and build the Glasgow Haskell Compiler (GHC) and GUM RTS. Next, GpH and Evaluation Strategies are briefly introduced, followed by a short discussion of parallelisation, profiling and optimisation. Subsequently, we illustrate how to make small incremental changes to the RTS, touch on debugging and conclude by presenting some open issues and future work directions. Basic knowledge of Linux tools, C and Haskell is assumed. 1 Setting Up the Development Environment To benefit from running GpH [38] programs using the distributed GUM [37] RTS [5] instead of the standard GHC RTS, we need a distributed target platform such as a cluster of multi-cores or a heterogeneous system such as a node comprising at least one host CPU (e.g. Intel Xeon) and at least one co-processor (e.g. Xeon Phi) which communicate using message passing (for instance using Virtual TCP/IP over PCIe). Adding and exploiting GPUs and accelerator cores to this configuration remains for future work. We assume a standard Linux environment (in our case CentOS 6.7; use $ uname -a and $ cat /etc/redhat-release to check). Below we discuss some common command line utilities that should be available by default and several libraries that might need installation, notably PVM [32] or MPI [10] on which the messaging sub-system is built.