EDAN - Towards Understanding Memory Parallelism and Latency Sensitivity in HPC

Explore memory parallelism and latency sensitivity in high-performance computing through this 23-minute conference talk from the International Conference on Supercomputing (ICS'25). Learn about EDAN (Execution DAG Analyzer), a novel performance analysis tool that addresses the challenges of resource disaggregation in large-scale computing systems by analyzing memory access latency impacts. Discover how EDAN leverages runtime instruction traces to generate execution DAGs, enabling estimation of latency sensitivity in sequential programs and investigation of different hardware configurations. Understand the tool's capability to calculate theoretical performance bounds and provide insights into memory-level parallelism inherent to HPC applications. Examine practical applications through case studies of PolyBench, HPCG, and LULESH benchmarks that reveal characteristics of intrinsic memory-level parallelism and latency sensitivity. Gain insights into how this approach offers advantages over traditional custom hardware or cycle-accurate simulators by providing flexibility and efficiency in performance analysis for disaggregated computing environments.