Evaluation of OpenMP Dependent Tasks with the KASTORS Benchmark Suite

Abstract

International audienceThe recent introduction of task dependencies in the OpenMP specifi-cation provides new ways of synchronizing tasks. Application programmers can now describe the data a task will read as input and write as output, letting the runtime system resolve fine-grain dependencies between tasks to decide which task should execute next. Such an approach should scale better than the excessive global synchronization found in most OpenMP 3.0 applications. As promising as it looks however, any new feature needs proper evaluation to encourage applica-tion programmers to embrace it. This paper introduces the KASTORS benchmark suite designed to evaluate OpenMP tasks dependencies. We modified state-of-the-art OpenMP 3.0 benchmarks and data-flow parallel linear algebra kernels to make use of tasks dependencies. Learning from this experience, we propose extensions to the current OpenMP specification to improve the expressiveness of dependen-cies. We eventually evaluate both the GCC/libGOMP and the CLANG/libIOMP implementations of OpenMP 4.0 on our KASTORS suite, demonstrating the in-terest of task dependencies compared to taskwait-based approaches

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

“…By far, the most common codes found in evaluations of tools of this kind are matrix operations performed by blocks [8], [16], [36]. This is not surprising, as this class of algorithms can present very complex patterns of task dependencies that are both the largest challenge and the largest opportunity for performance improvement for these tools, particularly when compared to approaches that cannot support arbitrary patterns of dependencies.…”

“…Two of the limitations of this approach have been discussed in [36]. For example, this mechanism alone does not allow to use dependent tasks for reduction processes, forcing to use other OpenMP functionalities such as atomic operations or critical sections, or resort to other high level constructs such as section or for.…”

“…Despite the relevance of this approach, we have not found comparative studies of the existing practical alternatives of this kind beyond some performance comparisons [36]. For this reason this paper tackles this issue describing and comparing some of the high-level approaches available nowadays with a particular focus on their semantics and ease of use, our main aim being to help identify the best tool for each problem at hand.…”

Evaluating the Impact of OpenMP 4.0 Extensions on Relevant Parallel Workloads

“…The main improvement of the work presented in this paper is that we do not only use the tasking features available in OpenMP 4.0 but also suggest and evaluate new ones. In contrast, the mentioned KASTORS approach [15] suggests new features, different from the ones proposed in this paper, but does not evaluate them.…”

“…Similarly, the KASTORS suite [15] uses the OpenMP 4.0 task dependency constructs to extend the Cholesky and QR decompositions from the PLASMA library [11]. Also, the KASTORS suite provides a parallelized Poisson equation based kernel and extends the SparseLU and Strassen benchmarks from the Barcelona OpenMP Tasks Suite [8].…”