Task-to-accelerator mapping for predictable OpenMP applications

Abstract

OpenMP offers a high-level, portable programming model that simplifies the expression of parallelism and offloading in heterogeneous platforms, making it an ideal framework for managing task execution across CPUs and GPUs. However, as applications increasingly require predictability and multi-GPU execution, the lack of predictability in current OpenMP scheduling mechanisms poses significant challenges. Irregular workloads, variable data transfers, and dynamic resource contention often lead to inconsistent response times and reduced timing determinism. This unpredictability limits system efficiency and reliability, particularly in performance-critical and real-time domains. To overcome the challenge mentioned above, this paper proposes an efficient task-to-accelerator mapping (i.e., assigning tasks to GPU devices for execution) for multi-GPU OpenMP applications. The mechanism includes an offloading manager that operates a global queue and multiple local queues, and different heuristic algorithms that first allocate tasks to the most appropriate queues and then select the most suitable tasks from each queue. The proposed heuristics aim to balance queue loads and enhance the work-conserving property of the mapping process to minimize response time and response-time variability. The paper also presents a thorough evaluation using both synthetic graphs and graphs representative of real-world applications, showing that the proposed mapping method outperforms the other relevant mapping methods in terms of application response time and response-time variability.