Workshops Archive

A Heuristic Based Cache Miss Classifier


Workshop: 8th Workshop on Programming and Performance Visualization Tools (ProTools)

Authors: Mo'ath Qutaish, David Abramson, and Chao Jin (The University of Queensland); Bronis R. de Supinski (Lawrence Livermore National Laboratory); and Jeffrey K. Hollingsworth (University of Maryland)

Abstract: Cache misses can limit HPC applications, yet conventional profilers only identify where misses occur and not their root cause. Capacity pressure and cache-set conflicts often produce similar hotspots but generally require different optimizations. Conventional classification uses a fully associative shadow cache but introduces additional state and lookup work. We present an explainable classifier that combines five complementary signals. The signals are computed over bounded counter periods and combined on a signed capacity-conflict scale. The method uses bounded counters and a compact address-history table while diagnosing root causes. Across 12 program configurations and 48 data-object scopes, the classifier agreed in every tested case with the more frequent capacity or conflict label among the independently generated ground-truth misses in that scope. On the Himeno benchmark, CMSA required 2.32 times the runtime of baseline Cachegrind. These results demonstrate the potential of bounded signals for explainable cache diagnoses without shadow-cache labels during prediction.


Back to 8th Workshop on Programming and Performance Visualization Tools (ProTools) Archive Listing

Back to Full Workshop Archive Listing