Speaker
Description
As the slowdown of Dennard scaling has shifted performance gains toward exploiting parallelism on increasingly heterogeneous many-core architectures, task-based programming models have emerged as a promising approach for managing complex workloads. Despite their growing adoption, however, understanding how runtime behavior translates into application performance remains a significant challenge.
To better understand the performance of task-based applications, we investigate the key performance drivers using both an irregular and a regular workload. Using LU decomposition as an irregular case study, we develop an application-specific analytical model that accurately predicts the optimal task granularity, eliminating the need for costly auto-tuning. We validate the analysis across OmpSs-2, GCC OpenMP, and HPX on both x86 and Arm platforms, demonstrating consistent performance trends across programming models and architectures. We then examine the remaining discrepancies between modeled and measured performance, attributing them to runtime scheduling decisions, including task placement and overheads. Finally, using a 2D stencil benchmark analyzed under the Roofline model, we isolate these runtime effects in a regular workload and show that task overprovisioning and NUMA-aware execution are critical for achieving high performance on modern multi-NUMA systems. Together, these two case studies illustrate how analytical modeling and runtime characterization provide complementary insights into the performance of task-based applications.