Concorde: Fast and Accurate CPU Performance Modeling with Compositional Analytical-ML Fusion

TL;DR

Concorde addresses the trade-off between accuracy and speed in CPU performance modeling by fusing simple analytical per-resource bottleneck models with a lightweight ML predictor to produce CPI estimates in constant time $\mathcal{O}(1)$. It builds compact performance distributions from trace analysis and per-resource bounds, then uses a shallow MLP to map these features to CPI across a large design space, enabling rapid design-space exploration and large-scale attribution. The approach achieves about $2\%$ average CPI error on unseen data, with massive speedups over cycle-level simulators (up to $7$ orders of magnitude for 1B-instruction programs) and demonstrates fine-grained Shapley-value based attribution across ARM N1-like cores. Concorde’s results suggest the value of compositional analytical-ML modeling for scalable architectural exploration, while acknowledging tail errors, data requirements, and potential extensions to more complex, multi-core or accelerator contexts.

Abstract

Cycle-level simulators such as gem5 are widely used in microarchitecture design, but they are prohibitively slow for large-scale design space explorations. We present Concorde, a new methodology for learning fast and accurate performance models of microarchitectures. Unlike existing simulators and learning approaches that emulate each instruction, Concorde predicts the behavior of a program based on compact performance distributions that capture the impact of different microarchitectural components. It derives these performance distributions using simple analytical models that estimate bounds on performance induced by each microarchitectural component, providing a simple yet rich representation of a program's performance characteristics across a large space of microarchitectural parameters. Experiments show that Concorde is more than five orders of magnitude faster than a reference cycle-level simulator, with about 2% average Cycles-Per-Instruction (CPI) prediction error across a range of SPEC, open-source, and proprietary benchmarks. This enables rapid design-space exploration and performance sensitivity analyses that are currently infeasible, e.g., in about an hour, we conducted a first-of-its-kind fine-grained performance attribution to different microarchitectural components across a diverse set of programs, requiring nearly 150 million CPI evaluations.

Figures (17)

• Figure 1: Per-resource analytical modeling produces a rich performance characterization of a program.
• Figure 2: Concorde's compositional analytical-ML structure
• Figure 3: Design overview
• Figure 4: Average test/train overlap across benchmarks
• Figure 5: Scatterplot of Concorde's CPI prediction error vs. the CPI for unseen (test) pairs of 100k-instruction regions and microarchitectural parameters. The plots on the sides show the distributions of CPI and prediction error. The average error is 2%, with only 2.5% of samples having larger than 10% error.