Adaptive and Architecture-Independent Task Granularity for Recursive Applications

Abstract

In the last few decades, modern applications have become larger and more complex. Among the users of these applications, the need to simplify the process of identifying units of work increased as well. With the approach of tasking models, this want has been satisfied. These models make scheduling units of work much more user-friendly. However, with the arrival of tasking models, came granularity management. Discovering an application's optimal granularity is a frequent and sometimes challenging task for a wide range of recursive algorithms. Often, finding the optimal granularity will cause a substantial increase in performance. With that in mind, the quest for optimality is no easy task. Many aspects have to be considered that are directly related to lack or excess of parallelism in applications. There is no general solution as the optimal granularity depends on both algorithm and system characteristics. One commonly used method to find an optimal granularity consists in experimentally tuning an application with different granularities until an optimal is found. This paper proposes several heuristics which, combined with the appropriate monitoring techniques, allow a runtime system to automatically tune the granularity of recursive applications. The solution is independent of the architecture, execution environment or application being tested. A reference implementation in OmpSs-a task-parallel programming model-shows the programmability, ease of use and competitive performance of the proposed solution. Results show that the proposed solution is able to achieve, for any scenario, at least 75% of the performance of optimally tuned applications.

Combining Dynamic Concurrency Throttling with Voltage and Frequency Scaling on Task-based Programming Models

“…To categorize tasks, we extend our monitoring and prediction infrastructure [6] that collects timing and hardware counter metrics at run-time, and infers these metrics for future tasks. Inferring predictions using basic metric averages may produce adverse effects.…”

Enhancing Resource Management through Prediction-based Policies

“…For instance, one of the inputs may be too large to fit within the same cache hierarchy level. Thus, to solve this, we use the cost clause, already proposed in previous works [15]. This clause specifies, in a rough way, the computational weight of a task.…”

Worksharing Tasks: An Efficient Way to Exploit Irregular and Fine-Grained Loop Parallelism

“…Consequently, a given problem can be solved either by using many fine-grained tasks or a few coarsegrained ones. Thus, finding an adequate granularity is a key point to optimally exploit resources when using tasks [9], alleviating the aforementioned overheads, but still creating enough parallelism to maximize resource utilization.…”