paperbot · PL 论文追踪

全部论文 待读 收藏

P2 Performance Predictability in Heterogeneous Memory

异构内存中的性能可预测性
ASPLOS Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 22026匹配度 1/10 · 该论文聚焦异构内存(DRAM+CXL)的性能预测与运行时策略,属于体系结构与内存管理方向,不涉及静态分析、LLM/Agent、浮点程序分析或贝叶斯程序分析。其分析对象是硬件性能计数器与内存层级行为,与用户的研究画像没有直接交集。引用 2 inbox
Jinshu Liu, Hanchen Xu, Daniel S. Berger, Marcos K. Aguilera, Huaicheng Li
crossref: 2026-09-12dblp: 2026-09-12

AI 速览

一句话:提出 CAMP 框架,用 12 个性能计数器解析式地将 CXL 引起的减速分解为三个正交分量,实现 10% 绝对误差内 91–97% 的预测准确率。

异构内存将 DRAM 与 CXL 组合使用,性能波动大,但现有指标与实际减速相关性弱。本文提出 CAMP:核心洞察是一次 DRAM 运行(带宽受限负载再加一次 CXL 运行)即可暴露 CXL 延迟转化为额外处理器停顿周期的因果微架构压力点。CAMP 用 12 个性能计数器将减速解析分解为三个正交分量——需求读、缓存/预取与写,并给出软件加权交织的闭式模型,预测不同 DRAM–CXL 比例下的性能。在 NUMA 与三种 CXL 设备上的 265 个负载中,10% 绝对误差内预测准确率达 91–97%;基于该模型的 Best-shot 交织与同置负载放置策略,相较现有分层与同置方案分别提升最多 21% 与 23%。

问题
DRAM 与 CXL 组成的异构内存性能高度可变,而现有指标与真实减速相关性弱,导致既难以预测性能,也难以据此指导内存分层与负载放置决策。
方法
CAMP 通过一次 DRAM 运行(带宽受限负载再加一次 CXL 运行)采集 12 个性能计数器信号,解析式地将 CXL 引发的减速分解为需求读、缓存/预取和写三个正交分量,并为软件加权交织建立跨 DRAM–CXL 比例的闭式预测模型。
评估
在 NUMA 平台与三种 CXL 设备上对 265 个负载评估预测准确率(10% 绝对误差内达 91–97%),并基于模型实现 Best-shot 交织与同置负载放置策略,与现有分层及同置方案对比端到端性能提升。
原文摘要(Abstract)

Heterogeneous memory combining DRAM and CXL exhibits variable performance, yet existing metrics correlate weakly with actual slowdown. We present CAMP, a principled framework for predicting CXL-induced slowdown. Our key insight is that a DRAM run (plus a CXL run for bandwidth-bound workloads) exposes the causal microarchitectural pressure points where CXL latency translates into additional processor stall cycles. CAMP captures these signals using 12 performance counters to analytically decompose slowdown into three orthogonal components: demand reads, cache/prefetching, and stores. CAMP also introduces a closed-form model for software-based weighted interleaving that predicts performance across DRAM--CXL ratios. Across 265 workloads on NUMA and three CXL devices, CAMP achieves 91--97% prediction accuracy within 10% absolute error. We demonstrate that these models enable practical system policies, including ''Best-shot'' interleaving and colocated workload placement, improving performance by up to 21% and 23% over existing tiering and colocation approaches.

链接与引用

DOI 原文 · PDF(开放获取) · DBLP

BibTeX
@inproceedings{LiuXBAL26,
  title = {Performance Predictability in Heterogeneous Memory},
  author = {Jinshu Liu and Hanchen Xu and Daniel S. Berger and Marcos K. Aguilera and Huaicheng Li},
  booktitle = {ASPLOS},
  year = {2026},
  doi = {10.1145/3779212.3790201}
}