改善Linux網路儲存系統上Samba伺服器檔案寫入之效能

Performance Optimization of the Samba write service on Linux-based Network-Attached Storage Systems

黃昶竣(Chang-Chun Huang)

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(2008年)

https://doi.org/10.6342/NTU.2008.02263

摘要

隨著網路技術的進步，連網式儲存裝置也變得越來越普遍，特別於Network-Attached Storage（NAS）更為普遍，隨著NAS的普遍，NAS的價格與效能也越來越受重視，另外隨著網路頻寬的增加，使得較低階的處理器沒辦法負擔網路協定的處理以及資料的搬移。為了解決這些問題，有研究提出Remote Direct Memory Access（RDMA）及 TCP Offload Engine（TOE）等技術來解決這些問題。然而這些解決方案必須添加額外的硬體支援，還需要軟體的配合才能有效率地減輕CPU的負擔。
本論文利用效能追蹤量測工具及技術，找出NAS上所需改善的效能瓶頸，並利用追蹤到的資料協助軟硬體整合的工作。我們成功地將Samba NAS server處理資料的主要功能遷移到Linux作業系統核心（kernel），以減輕對CPU在資料搬移的負擔。由於軟體架構的改變，我們得到1.01~1.31的寫檔效能改進。未來搭配TOE之後，對於大檔案的寫入，預計可達到2.62的效能增進。

關鍵字

網路儲存系統 ； TCP卸載引擎

並列摘要

As the Ethernet performance increases, network-enabled storage solution becomes popular, especially for the Network Attached Storage （NAS） appliances. Some low-end NAS servers come with affordable prices and thus, the trade between performance and cost in such low-end NAS servers become an important issue. In fact, while processing the I/O requests from the client machine, the control processor in the low-cost NAS server spends most of the time handling data copy operations and network protocols. To boost the performance, software （zero-copy） and hardware （TCP Offload Engine） solutions are provided to remove data copy operations and handle network protocols, respectively. However, few attempts are made to discuss this hardware-software codesign issue about the software effort made to take advantage from the offload engines.
In this thesis, we tackle the hardware-software codesign issue by presenting a throughout performance study of target system. We first diagnose the performance of target system. The result shows that data copy operations and data processing in network protocol stack are two major performance bottleneck. For the software part, we migrate the Samba engine to the kernel and remove unnecessary data copy operations. For the hardware part, we predict the performance after adding a offload engine to handle data processing in network protocol stack. We show that an 1.01~1.31 speedup is achieved for the software part compared to original code. And, 1.51~2.62 performance improvement is enhanced if the network protocol process can be offload to a dedicated hardware.

並列關鍵字

Samba ； NAS ； zero-copy ； TOE

參考文獻

[19] Wu, Z.-Z., Chen, H.-C., and Huang, C.-M., "The 10gbit Hba Hardware Design for Iwarp Offloading Engine", CCL TECHNICAL JOURNAL, 2005.

[5] Dalessandro, D., Wyckoff, P., and Montry, G., "Initial Performance Evaluation of the Neteffect 10 Gigabit Iwarp Adapter", in Cluster Computing, 2006 IEEE International Conference, 2006, pp. 1-7.

[6] Balaji, P., Jin, H.W., Vaidyanathan, K., and Panda, D.K., "Supporting Iwarp Compatibility and Features for Regular Network Adapters", in Cluster Computing, 2005. IEEE International, 2005, pp. 1-10.

[8] Engel, J., Meneskie, J., and Kocak, T., "Performance Analysis of Network Protocol Offload in a Simulation Environment", in Atlantic Coast Marketing SE, 2006, pp. 762-763.

[9] Halvorsen, P., Jorde, E., Skevik, K.A., Goebel, V., and Plagemann, T., "Performance Tradeoffs for Static Allocation of Zero-Copy Buffers", in Proceedings of 28th Euromicro Conference, 2002, pp. 138-143.