The Memory Behavior of Data Structures in C SPEC CPU2000 Benchmarks

As a result of application and computer system design trends, the memory system continues to exert a dominant influence on program performance. While understanding how complex applications use complex computer systems is important to hardware, software, and benchmark designers, this is a difficult task. This paper demonstrates a system called DTrack that provides deeper insight into how an application uses the memory system. DTrack classifies memory accesses on a per data structure basis, enabling analysis of the regularity and locality of these program components. Applying the DTrack methodology to 12 C SPEC CPU2000 benchmarks, we demonstrate that this classification reveals data structure interactions that remain obscured with traditional aggregation-based analysis methods. Our characterization reveals the degree of diversity in memory behavior among the benchmarks within this suite, and we discuss how these insights can be used by system and application designers to achieve high performance through memory system and compiler improvements.