This is an experimental package for Julia programmers to implement parallel computations aimed at exploiting multi-level parallelism hierarchies. The current supported levels are multicluster, cluster, and multicore, with plans to include the support for accelerators in the near future.
It is still a proof-of-concept prototype for the concept of a multiscale parallel component model called μHash, in which parallel computing concerns at different levels of the parallelism hierarchy of a parallel computing platform are encapsulated in distinct components.
In a parallel programming language/artifact fully based on μHash, a particular implementation of a parallel computation that exploit multiple levels of the parallelism hierarchy of a parallel computer naturally comprises a set of components. They encapsulate the concerns behind the implementation of the computation for each level, separately. By consequence of such a modularity feature, this language/artifact itself may give support to the best programming abstractions, constructs, and primitives for exploiting efficiently the parallelism at each level, instead of relying on external artifacts (e.g., third-party libraries and frameworks).
In the current implementation of Hash.jl, using the metaprogramming features of Julia to avoid modifying its compiler or runtime system, components are still compile-time entities (static), and Julia programmers must use the existing programming artifacts of the Julia ecosystem (e.g. Distributed.jl, MPI.jl, FLoops.jl, @threads, @spawn, etc) in order to program for each parallelism level.
Hash.hl depends on another experimental package, called MulticlusterManager.jl in order to enable parallelism at the multicluster level.
Please contact us if you are interested in using this package or contributing to its development.