Faithful performance prediction of a dynamic task‐based runtime system for heterogeneous multi‐core architectures

Abstract

SUMMARYMulti-core architectures comprising several graphics processing units (GPUs) have become mainstream in the field of high-performance computing. However, obtaining the maximum performance of such heterogeneous machines is challenging as it requires to carefully off-load computations and manage data movements between the different processing units. The most promising and successful approaches so far build on task-based runtimes that abstract the machine and rely on opportunistic scheduling algorithms. As a consequence, the problem gets shifted to choosing the task granularity, task graph structure, and optimizing the scheduling strategies. Trying different combinations of these different alternatives is also itself a challenge. Indeed, obtaining accurate measurements requires reserving the target system for the whole duration of experiments. Furthermore, observations are limited to the few available systems at hand and may be difficult to generalize. In this article, we show how we crafted a coarse-grain hybrid simulation/emulation of StarPU, a dynamic runtime for hybrid architectures, over SimGrid, a versatile simulator of distributed systems. This approach allows to obtain performance predictions of classical dense linear algebra kernels accurate within a few percents and in a matter of seconds, which allows both runtime and application designers to quickly decide which optimization to enable or whether it is worth investing in higher-end graphics processing units or not. Additionally, it allows to conduct robust and extensive scheduling studies in a controlled environment whose characteristics are very close to real platforms while having reproducible behavior.

Performance modeling of a geophysics application to accelerate over‐decomposition parameter tuning through simulation

“…Our methodology is configured with a faithful model of the computation and communication signatures of Ondes3D for representative workloads on scale. All performance modeling is carried out with SimGrid's SMPI emulation and trace replay mechanisms, which have received extensive validation efforts . We show that our methodology is faithful, in terms of total makespan as well as from the load balancing perspective observed in real AMPI runs.…”

Exploring Scheduling Effects on Task Performance with TaskInsight

“…Stanisic et al [9] simulate tasks' execution in order to isolate the scheduler's effect on performance from tasks' unpredictable behavior. In our work we execute the entire application on real hardware and we characterize the interaction between the scheduler and the tasks, which triggers online scheduling decisions and performance variation.…”

“…Typical approaches to optimizing scheduling algorithms consist of either providing an interactive visualization of the execution trace [1,5] or simulating the tasks execution to evaluate the overall scheduling policy in a controlled environment [4,9]. A developer then has to analyze the resulting profiling information and deduce if the scheduler behaves as expected, and qualitatively compare different schedulers.…”

Special Issue: Euro‐Par 2014

“…Eventually, four papers were accepted for publication. They cover the following four Euro-Par topics: Performance Prediction and Evaluation, Distributed Systems and Algorithms, Theory and Algorithms for Parallel Computation and High-Performance and Scientific Applications.Topic 2 on Performance Prediction and Evaluation contributes the paper Performance prediction of dynamic task-based runtime system for heterogeneous multi-core architectures authored by Luka Stanisic, Samuel Thibault, Arnaud Legrand, Brice Videau and Jean-François Méhaut [1]. They address the challenge of deciding, with high accuracy and low effort, what computations to offload onto accelerators on a heterogeneous execution platform.…”

“…Topic 2 on Performance Prediction and Evaluation contributes the paper Performance prediction of dynamic task-based runtime system for heterogeneous multi-core architectures authored by Luka Stanisic, Samuel Thibault, Arnaud Legrand, Brice Videau and Jean-François Méhaut [1]. They address the challenge of deciding, with high accuracy and low effort, what computations to offload onto accelerators on a heterogeneous execution platform.…”