Fine-grained data usage analysis by access sampling: seeing the data that is not there

Estimating active data usage is a basic problem in memory system analysis, management and optimization. Fine-grained usage analysis is costly because it requires monitoring data access. This paper presents efficient fine-grained analysis through access sampling. By taking random samples at some frequency ratio, e.g. 1% of cache misses, it infers the size of the other data accessed in the rest of the trace. Since the analysis deduces the total amount of data accessed by inspecting a subset of accesses, it is seeing the data that is not there. The paper presents the analysis and its evaluation using 8 program traces. The error of data-size prediction is 33% at 1% sampling and 6% at 10% sampling. The new technique is significantly more accurate than two previous models. One is based on skewed distributions, i.e. the "80-20" law. The other is the well-known Good-Turing frequency estimation.