| 王鹿鸣,卢天越,陈明宇.远端内存访问性能优化关键技术综述[J].高技术通讯(中文),2025,35(6):565~578 |
| 远端内存访问性能优化关键技术综述 |
| Survey of key technologies for optimizing far memory access performance |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 06. 001 |
| 中文关键词: 远端内存；内存级并行；延迟容忍 |
| 英文关键词: far memory, memory-level parallelism, latency tolerance |
| 基金项目: |
| 作者 | 单位 | | 王鹿鸣 | (中国科学院计算技术研究所北京 100190)
(中国科学院大学北京 100049) | | 卢天越 | (中国科学院计算技术研究所北京 100190)
(中国科学院大学北京 100049) | | 陈明宇 | (中国科学院计算技术研究所北京 100190)
(中国科学院大学北京 100049) |
|
| 摘要点击次数: 13 |
| 全文下载次数: 12 |
| 中文摘要: |
| 在数据中心，应用对于内存的需求越来越大。一系列提供更大容量但访问延迟也更高的远端内存技术被引入到数据中心中。然而，其较长的访存延迟和高波动性对于传统的处理器体系结构构成了全新的挑战。本文系统性地总结了远端内存对传统处理器设计带来的挑战，并综述了近年来的优化访存延迟的关键技术及其在远端内存应用上的优缺点。本文归纳指出，解决远端内存访存问题需要采用软硬件结合的方案。硬件提供足够的资源维护并发的访存请求，软件辅助参与指令调度；在支持大量并发访存请求的基础上，需要平衡专用性与通用性，尽量不牺牲通用处理器的性能。此外，还需考虑优化分支误预测的处理机制，减少猜测执行失败回滚对系统性能造成的影响。 |
| 英文摘要: |
| In data centers, memory demands for applications are increasing rapidly. A variety of far memory technologies that provide larger capacity but also higher access latency have been introduced into data centers. However, the longer access latency with higher variability of far memory poses challenges for conventional computer architectures. This paper systematically summarizes the challenges that far memory brings to conventional processor microarchitecture designs. Additionally, this paper reviews the key technologies in recent years for latency-tolerancing, summarizing their pros and cons when applied to far memory scenarios.This paper concludes that solving the issues of accessing far memory requires a cooperative software-hardware solution. Hardware provides sufficient resources to maintain outstanding memory requests, while software assists in dynamic instruction scheduling. Additionally, a balance must be struck between specialization and generality of the solution. Furthermore, mechanisms for optimizing branch misprediction handling need to be considered to reduce the performance impact of mis-speculation. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
