Greedy Adaptive Task Splitting with Prediction and Placement for Heterogeneous Systems
Résumé
Using task-based programming models enables the efficient exploitation of heterogeneous parallel systems. However, these models are often restricted to task graphs whose granularity is fixed at submission time, making it difficult to determine the optimal task size. In heterogeneous systems, for example, GPUs require coarse-grained tasks to achieve high efficiency, while executing such tasks on CPU cores is prohibitive due to their long duration.
Recursive tasks address this issue by allowing multiple granularities within the same graph, dynamically splitting tasks into subDAGs at runtime through the Splitter. The splitting decision can be guided by a Linear Programming (LP) strategy, as introduced in [1].
In this paper, we introduce GASPP, a new task splitting strategy that overcomes the limitations of the LP strategy. Our greedy approach leverages runtime predictions and integrates scheduling constraints to ensure effective CPU utilization. We then present an experimental evaluation demonstrating that our new strategy improves performance across a range of dense linear algebra algorithms.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| licence |
