Tools-MAAP: A Metadata-Aware Application Profiling Framework for Memory System Characterization

Abstract

Memory-system benchmarking combines dynamic instrumentation, hardware counters, and architectural simulation, yet the resulting measurements are often separated from the compiler, input, platform, profiler, and runtime context that produced them. This limits reproducibility and makes observations collected with different methodologies difficult to organize and compare. We present MAAP, a metadata-aware workload-characterization framework that unifies DynamoRIO, Sniper, and perf for CPUs with NVBit and Nsight Compute for NVIDIA GPUs. MAAP captures five categories of execution metadata from build files, system interfaces, runtime configuration, and profiler outputs, maintains them in a persistent metadata object, and attaches its identifier to every aggregate, trace, time-series, and architectural record. Metadata collection adds roughly 10–30KB per session with sub-second latency. Using this methodology, we release more than 2,400 metadata-paired observations across SPEC CPU2017, SPEC CPU2026, and PolyBenchGPU, enabling researchers to construct reproducible benchmark subsets and explore compiler, input, platform, profiler, and GPU configuration settings. MAAP, the dataset, and documentation are released as open source at https://github.com/TuftsECS/MemSysExplorer/tree/main/apps.