CASA: 정적 접근 패턴 기반 캐시 분석 프레임워크의 설계 및 구현

초록

Cache behavior has a significant impact on program performance, and analyzing cache usage before execution is important for identifying performance bottlenecks and deriving optimization directions. This paper presents the design and implementation of CASA (Cache Analysis Framework Based on Static Access Patterns), which analyzes cache behavior without executing the target program using static access patterns extracted from LLVM IR. CASA transforms static access patterns into a memory access sequence and analyzes hit/miss behavior on a YAML-configured cache model. It also classifies cache misses using a fully associative comparison cache and aggregates the results by memory object and Load/Store operation. In an evaluation with PolyBench benchmarks, CASA showed mean absolute percentage errors of 0.7% and 0.1% for L1 and LL cache miss counts compared with Cachegrind, demonstrating results consistent with an existing dynamic profiling tool.

키워드