以遠端記憶體優化虛擬記憶體及檔案系統快取

Enhancing Virtual Memory and Filesystem Cache with Remote Memories

涂世昱(Shih-Yu Tu)

指導教授 ： 洪士灝

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202100453

摘要

在以大型的資料為主的高效能運算應用中,例如基因分析,資料儲存裝置時常成為效能的瓶頸。為了有效增進儲存裝置資料交換的效率,在本篇論文當中我們利用遠端記憶體存取的技術,將遠端機器的記憶體作為近端機器儲存裝置的快取,在兩台機器之間以高速的網路界面控制器作為資料交換的媒介。相較於傳統以固態硬碟作為快取裝置的作法,我們所提出的方法有更高的頻寬與可靠度。我們針對了不同的資料存取的模式個別進行簡單的效能評測,在循序讀取方面,相較於將資料存放在固態硬碟的方式,我們提出的方法可以達到 3.5 倍的加速。另外我們也以實際基因分析的案例 (bigsnpr) 做效能測試,實驗結果顯示有 2 倍的效能加速。

關鍵字

遠端記憶體直接存取 ； 資料交換 ； 快取 ； 資料儲存 ； 虛擬記憶盤

並列摘要

For high-performance computing applications with big data, such as genomic analysis, the storage devices often become the performance bottleneck. In order to improve the performance of storage devices, we propose to borrow memory from a remote server via a high bandwidth network to serve as a caching device for the local machine’ s storage device. Our proposed method offers higher bandwidth and reliability compared to the traditional practice of using SSD as a caching device, and our implementation takes advantage of the remote direct memory access (RDMA) capability supported by the network adapter to reduce the overhead of accessing the proposed cache device. We have carried out simple micro-benchmarks as well as a application case study to demonstrate the benefits of the proposed method. In a sequential read test, compared to the traditional method of storing data on an SSD, the proposed method achieved a speedup of 3.5x. As to the real-world workload for running genomic analysis with bigsnpr, our proposed method provided 2x speedup against the traditional practice.

並列關鍵字

RDMA ； I/O ； cache ； storage ； ramdisk

參考文獻

[1] 3d xpoint technology. https://www.micron.com/products/advanced-solutions/3d-xpoint-technology.

[2] Nvme over fabrics. https://nvmexpress.org/wp-content/uploads/NVMe_Over_Fabrics.pdf.

[3] Page frame reclamation. https://www.kernel.org/doc/gorman/html/understand/understand013.html.

[4] B. Abali, R. Eickemeyer, H. Franke, C.-S. Li, and M. Taubenblatt. Disaggregated and optically interconnected memory: when will it be cost effective? 03 2015.

[5] M. K. Aguilera, N. Amit, I. Calciu, X. Deguillard, J. Gandhi, P. Subrahmanyam, L. Suresh, K. Tati, R. Venkatasubramanian, and M. Wei. Remote memory in the age of fast networks. In Proceedings of the 2017 Symposium on Cloud Computing, SoCC ’17, page 121– 127, New York, NY, USA, 2017. Association for Computing Machinery.