位置推广策略优化最后一级缓存性能

需积分: 10 99 浏览量更新于2024-08-26 收藏 1.27MB PDF 举报

本文研究了"最后一级缓存的基于位置的促销政策"（Position-Based Promotion Policy, PPP），针对内存密集型工作负载优化缓存性能。在处理处理器与主内存之间长时间延迟的问题时，传统的最少使用（Least Recently Used, LRU）替换策略依赖于最近访问频率来决定块的去留。然而，这种方法可能无法准确预测块的重用情况，导致零重用块占用时间过多，降低了高局部性块的缓存效率。近期的研究指出，改变缓存插入时的重用预测策略，如通过更精确地预测传入块的重用间隔（Re-Reference Interval Prediction, RRIP），能够提高缓存性能。与RRIP策略相结合，PPP提出了一种新颖的方法，它根据块在RRIP链中的位置来预测重用概率。这样做的优势在于，它不仅考虑了块的新近度（即访问频率），还考虑了块的定位信息，从而减少了由于频繁重新引用但不再使用的块导致的性能下降。 PPP的优点在于对硬件改动较小，只需对RRIP算法进行微调。实验结果显示，PPP相较于原始RRIP，能实现大约0.74%的加速性能提升，相比于LRU策略，其性能分别提高了3.2%和2.4%。这意味着PPP在保持高性能的同时，优化了缓存的利用效率，对于现代数据中心和计算密集型应用来说，具有显著的实际意义。总结来说，这篇研究深入探讨了如何通过位置信息驱动的促销策略来改进最后级缓存的管理，提供了一种创新的解决方案，旨在改善内存密集型任务的执行速度，挑战了传统的缓存替换策略。这对于优化计算机系统性能、减少延迟并提升整体计算效率具有重要价值。

RESEARCH ARTICLE

Printed in the United States of America

Journal of

Computational and Theoretical Nanoscience

Vol. 12, 1–10, 2015

Position-Based Promotion Policy for the

Last Level Cache

Manman Peng

∗

, Bin Yu, and Tingting Zhu

College of Computer Science and Electronic Engineering, Hunan University, Changsha, China

The last level cache plays an impor tant role in mitigating long latencies between processor and main

memory. Recent studies have shown that changing the re-reference (or reuse) prediction on cache

insertions can signiﬁcantly improve cache performance for memory-intensive workloads. Unlike

least-recently-used (LRU) replacement policy, these policies make the prediction of the incoming

blocks more correctly and hence reduce the amount of time which the zero-reused blocks occupy.

As a result, the high locality blocks get more opportunity to reside in the cache. However these

policies make the same prediction on a cache hit. That is, on any cache hit, the block is promoted to

the head of the ordered chain. This can potentially degrade cache performance when a cache block

is re-referenced and then never reused again. We show that simple changes to the promotion policy

can signiﬁcantly reduce cache misses for memor y-intensive workloads. We propose Position-based

Promotion Policy (PPP) that predicts the re-reference interval of the reuse block based on its posi-

tion in the Re-Reference Interval Prediction (RRIP) chain. When combined with RRIP, PPP takes

both recency and frequency information into consideration at the same time. PPP requires minor

hardware modiﬁcation on the RRIP. Our evaluations show that it achieves a speedup of 0.74% over

the original RRIP, and they outperfor m LRU by 3.2% and 2.4% respectively.

Keywords: Last Level Cache, Insertion Policy, Promotion Policy, Less Reused Block, Locality.

1. INTRODUCTION

The growing performance gap between processors and

memory has long been a primary bottleneck fo r hig h-end

processors. The last level cache (LLC) in cache hierar-

chy plays an important role in mitigating the processor-

memory gap by exp loiting temporal and spatial locality.

For many years, the cache has been managed by the least-

recently-used (LRU) replacement policy and its approx-

imations. Recent studies, however, have shown that the

LRU rep lacement policy can still be improved for appli-

cations which have a working set larger than the cache

or have mixed re-references pattern.

1, 2

As a result, some

of the studies proposed novel ideas to address the prob-

lem by simply modifying the cache insertion policy of

LRU replacement policy. This paper further squeezes the

room for cache performance improvement by addressing

the limits of the prior work.

Re-Reference Interval Prediction (RRIP),

a recently-

pro-posed framework, have altered the description of the

LRU chain. Rather than representing recency, it treats

∗

Author to whom correspondence should be addressed.

LRU chain as RRIP chain that represents the predicted re-

reference order of cache blocks. Cache block at the head

or the tail of the RRIP chain is predicted to have a near-

immediate or distant re-reference interval respectively.

Using the RRIP framework, the commonly-used LRU

replacement policy predicts all incoming blocks to have a

near-immediate re-reference interval and makes the same

prediction on a cache hit. Recent studies

1, 2

have shown

that always predicting a near-immediate re-reference inter-

val is not robust across all access patterns which may

contain either near-immediate o r distant re-references or

both. These access patterns was usually found in the work-

loads that have a working set larger than the cache or

have frequent bursts of references to non temporal data.

For such workloads, LRU blindly inserts all the blocks at

the head of the RRIP chain. These blocks which have a

distant re-reference interval travel from the most-recently-

used (MRU) position to the LRU position without receiv-

ing any cache hits and pollute the active block in the cache,

resulting inefﬁcient use of cache space.

Dynamic Re-Reference Interval Prediction (DRRIP)

addresses this problem by dynamically choosing between

the two techniques, Static Re-Reference Interval Prediction

J. Comput. Theor. Nanosci. 2015, Vol. 12, No. 10 1546-1955/2015/12/001/010 doi:10.1166/jctn.2015.4264 1

下载后可阅读完整内容，剩余9页未读，立即下载

weixin_38586118

粉丝: 6
资源: 922

位置推广策略优化最后一级缓存性能

大促抗住零点洪峰-缓存架构体系课件

高并发秒杀系统中的缓存优化策略

【数据库缓存策略】：利用django.db.connection实现高效缓存，提升数据处理速度

【Django缓存最佳实践】：分享基于django.core.cache.backends.base的最佳实践技巧

Django日期计算优化：缓存策略减少重复计算成本

Python Django缓存策略：提升网站性能的实用技巧

C#缓存高级技巧：3大诀窍打造极致性能应用

Django Admin缓存策略：2种方法优化后台响应速度

Python CGI缓存策略：提升Web应用性能的实战技巧

Java企业应用中的缓存策略：性能提升的关键技术揭秘

最新资源