| PM (Parallel Models) is a new parallel programming language specifically designed for
writing environmental and geophysical models. The language is intended to enable
implementers to concentrate on the science behind the model rather than the details of
running on parallel hardware. At the same time PM leaves the programmer in control – all
parallelisation is explicit and the parallel structure of any given program may be deduced
directly from the code.
This paper describes a PM implementation based on the Message Passing Interface (MPI)
and Open Multi-Processing (OpenMP) standards, looking at issues involved with translating
the PM parallelisation model to MPI/OpenMP protocols and considering performance in
terms of the competing factors of finer-grained parallelisation and increased communication
overhead. In order to maximise portability, the implementation stays within the
MPI 1.3 standard as much as possible, with MPI-2 MPI-IO file handling the only
significant exception. Moreover, it does not assume a thread-safe implementation of
MPI.
PM adopts a two-tier abstract representation of parallel hardware. A PM processor is a
conceptual unit capable of efficiently executing a set of language tasks, with a complete
parallel system consisting of an abstract N-dimensional array of such processors. PM
processors may map to single cores executing tasks using cooperative multi-tasking, to
multiple cores or even to separate processing nodes, efficiently sharing tasks using algorithms
such as work stealing. While tasks may move between hardware elements within a PM
processor, they may not move between processors without specific programmer intervention.
Tasks are assigned to processors using a nested parallelism approach, building on
ideas from Reyes et al. (2009). The main program owns all available processors.
When the program enters a parallel statement then either processors are divided out
among the newly generated tasks (number of new tasks < number of processors) or
tasks are divided out among the available processors (number of tasks > number of
processors). Nested parallel statements may further subdivide the processor set owned
by a given task. Tasks or processors are distributed evenly by default, but uneven
distributions are possible under programmer control. It is also possible to explicitly
enable child tasks to migrate within the processor set owned by their parent task,
reducing load unbalancing at the potential cost of increased inter-processor message
traffic.
PM incorporates some programming structures from the earlier MIST language presented
at a previous EGU General Assembly, while adopting a significantly different underlying
parallelisation model and type system. PM code is available at www.pm-lang.org under an
unrestrictive MIT license.
Reference
Ruymán Reyes, Antonio J. Dorta, Francisco Almeida, Francisco de Sande, 2009.
Automatic Hybrid MPI+OpenMP Code Generation with llc, Recent Advances in Parallel
Virtual Machine and Message Passing Interface, Lecture Notes in Computer Science Volume
5759, 185-195 |