QuaL(2)M : Learning Quantitative Performance of Latency-Sensitive Code

Quantitative performance predictions are more informative than qualitative. However, modeling of latency-sensitive code, with cost distributions of high variability and heavy tails, is extremely difficult. To date, quantitative prediction for such code has been limited to either special cases (e.g., best-case performance) or resource-intensive methods (e.g., simulation). We present QuaL(2)M, a method for learning quantitative performance of latency-sensitive CPU code. We collect high resolution data (superblock, i.e., short instruction sequence), with challenging cost distributions, from several applications over a range of inputs and times. For each superblock, QuaL(2)M predicts both expected and degraded performance in cycles. QuaL(2)M distinguishes superblock behavior by combining lightweight telemetry from performance monitoring units and readily obtainable compiler execution models. Compared to two state-of-the art methods, on our largest dataset, QuaL(2)M achieves an R-2 of 0.87 vs. 0.37 and 0.39. The trained models can be used for online performance diagnosis and adaptation.