The rising on-node concurrency, combined with limited resources, makes it increasingly hard for a single application to exploit the computational power of a node. Co-Scheduling, i.e., the concurrent use of a single node by two or more complementary applications, can help mitigate this situation and achieve higher efficiency. In this paper, we propose an extension to HALadapt, a library for task-based programming, which leverages its dynamic profiling approach to provide concurrency throttling and combines it with its ability to coordinate execution across multiple runtime instances. Using a real-world example application, MLEM, co-scheduled with a compute-intensive synthetic workload stressgen, we show that the runtime system of HALadapt can efficiently coordinate multiple independent instances on a single node, leading to a performance improvement of up to 43% in the best case.     «

The rising on-node concurrency, combined with limited resources, makes it increasingly hard for a single application to exploit the computational power of a node. Co-Scheduling, i.e., the concurrent use of a single node by two or more complementary applications, can help mitigate this situation and achieve higher efficiency. In this paper, we propose an extension to HALadapt, a library for task-based programming, which leverages its dynamic profiling approach to provide concurrency throttling a...     »