一句话：通过三个工业产品的大规模研究发现多级 PGO 最高增益 17.04% 却也能带来 81.90% 的回退，并提出跨层安全应用指南 ML-PGO。
PGO 被广泛用于提升应用与操作系统内核性能，但已有工作都在单一层次上孤立地应用 PGO，忽略了内核与应用间的跨层交互：二者在生产环境中通常用不同工作负载独立优化，导致训练画像与最终协同部署的系统不匹配。本文在三个代表性工业产品上开展大规模多级 PGO 研究，在真实部署条件下评估独立优化的内核与应用组合后的整体效应。结果显示，多级 PGO 可带来最高 17.04% 的性能提升，但也可能造成最高 81.90% 的严重回退。通过性能计数器分析，作者将这些效应归因于画像假设不匹配所引发的微架构因素，如缓存竞争与分支预测错误，表明 PGO 收益在软件层次间并非可组合。基于此，作者提出 Multi-Level PGO（ML-PGO）这一实用指南，强调工作负载代表性与跨层感知对可靠性能提升的重要性。
Profile-guided optimization (PGO) is used to improve software performance in applications and operating system kernels. However, prior work applies PGO to a single layer in isolation, overlooking cross-layer interactions in systems, where kernel and application are optimized independently using different workloads. This practice is common in production environments and makes its impact on end-to-end performance important to understand. It introduces a mismatch between training and deployment, where profiling workloads may not reflect the final co-executed system. Despite its prevalence, the implications of such cross-layer mismatch remain poorly understood. In this paper, we present a large-scale industrial study of multi-level PGO across three representative products. We evaluate the combined effects of independently optimized kernel and application under realistic deployment conditions. Our results show that multi-level PGO can yield performance gains of up to 17.04%, but may lead to severe regressions of up to 81.90%. Through performance counter analysis, we attribute these effects to microarchitectural factors such as cache contention and branch misprediction arising from mismatched profiling assumptions. These findings show that PGO benefits are not necessarily compositional across software layers. Based on these insights, we propose Multi-Level PGO (ML-PGO), a practical guideline for applying PGO safely across system layers. Our work highlights the importance of workload representativeness and cross-layer awareness for achieving reliable performance improvements.
@inproceedings{paperbot12358,
title = {Beyond Single-Level PGO: Understanding and Harnessing Cross-Level Optimization Effects in Industrial Systems},
author = {Dong Liu and Yifan Cheng and Yunfeng Cui and Ting Chen and Jiachi Chen and Zhihong Xue and Chong Hu and Shengyu Cheng},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834484}
}