Dissecting on-node memory access performance: a semantic approach

摘要: Optimizing memory access is critical for performance and power efficiency. CPU manufacturers have developed sampling-based measurement units (PMUs) that report precise costs of accesses at specific addresses. However, this data too low-level to be meaningfully interpreted contains an excessive amount irrelevant or uninteresting information. We a method gather fine-grained objects regions code with low overhead attribute semantic information the sampled accesses. This provides context necessary more effectively interpret data. tool performs sampling attribution used discover diagnose problems in real-world applications. Our techniques provide useful insight into behaviour applications allow programmers understand ramifications key design decisions: domain decomposition, multi-threading, motion within distributed systems.