A Comparison of Task Parallel Frameworks based on Implicit Dependencies in Multi-core Environments

Abstract

Abstract-The larger flexibility that task parallelism offers with respect to data parallelism comes at the cost of a higher complexity due to the variety of tasks and the arbitrary patterns of dependences that they can exhibit. These dependencies should be expressed not only correctly, but optimally, i.e. avoiding over-constraints, in order to obtain the maximum performance from the underlying hardware. There have been many proposals to facilitate this non-trivial task, particularly within the scope of nowadays ubiquitous multi-core architectures. A very interesting family of solutions because of their large scope of application, ease of use and potential performance are those in which the user declares the dependences of each task, and lets the parallel programming framework figure out which are the concrete dependences that appear at runtime and schedule accordingly the parallel tasks. Nevertheless, as far as we know, there are no comparative studies of them that help users identify their relative advantages. In this paper we describe and evaluate four tools of this class discussing the strengths and weaknesses we have found in their use.

Easy Dataflow Programming in Clusters with UPC++ DepSpawn

“…The result is also a testament to the high degree of optimization of the standard mode. In fact, as one would expect, the kernel of UPC++ DepSpawn in standard mode is very much based on the runtime of DepSpawn, which showed to be on par with state of the art tools such as OpenMP in [10].…”

“…Its main responsibility is the scheduling of these tasks once DepSpawn informs TBB that they are ready for execution. Finally, it deserves to be mentioned that DepSpawn, which only supports shared-memory environments, has been compared in terms of performance and programmability to some of the most relevant alternatives in this field achieving satisfactory results [10].…”

“…A very effective mechanism to express this kind of algorithms is to rely on the implicit dependencies between the tasks that are given by the definition of their inputs and outputs. While this strategy is supported by several tools with very different nature [10], as far as we know, none of them is integrated and relies on the semantics of a PGAS © 2019 IEEE. This version of the article has been accepted for publication, after peer review.…”

High-performance dataflow computing in hybrid memory systems with UPC++ DepSpawn

“…This way the new library supports dataflow execution in a distributed memory system managed by UPC++. The library integrates also the shared memory runtime of the original DepSpawn, which was compared to state of the art approaches in [14] providing good results, thus enabling in addition task parallelism within each UPC++ process.…”

A software cache autotuning strategy for dataflow computing with UPC++ DepSpawn

“…This paradigm has repeatedly shown to largely improve the performance of critical mathematical algorithms, [9][10][11] therefore being of great practical interest. While several tools support this strategy, particularly in shared memory environments, 12 there were no PGAS alternative that supported it until the development of UPC++ DepSpawn. 13 This library enables dataflow computing on top of UPC++, 8 another library that implements and extends in C++ the Unified Parallel C (UPC) 3 language.…”