Memory Access Vectors: Improving Sampling Fidelity for CPU Performance Simulations

TL;DR

This paper introduces Memory Access Vectors (MAV) to enhance SimPoint sampling by capturing complex memory access patterns, significantly improving performance projection accuracy for array-indirect memory access benchmarks.

Contribution

The paper presents a novel combination of BBVs with MAVs, a microarchitecture independent technique, to improve sampling fidelity in CPU performance simulations.

Findings

Projection accuracy for 523.xalancbmk_r increased from 80% to 98%.

Memory Access Vectors effectively capture complex array-indirect memory behaviors.

Enhanced sampling method reduces simulation errors in large-scale CPU benchmarks.

Abstract

Accurate performance projection of large-scale benchmarks is essential for CPU architects to evaluate and optimize future processor designs. SimPoint sampling, which uses Basic Block Vectors (BBVs), is a widely adopted technique to reduce simulation time by selecting representative program phases. However, BBVs often fail to capture the behavior of applications with extensive array-indirect memory accesses, leading to inaccurate projections. In particular, the 523.xalancbmk_r benchmark exhibits complex data movement patterns that challenge traditional SimPoint methods. To address this, we propose enhancing SimPoint's BBV methodology by incorporating Memory Access Vectors (MAV), a microarchitecture independent technique that tracks functional memory access patterns. This combined approach significantly improves the projection accuracy of 523.xalancbmk_r on a 192-core system-on-chip, increasing it from 80% to 98%.