針對伺服器應用的系統和儲存設備的一鍵效能優化

Abstract

效能優化對於當代計算機系統至關重要，尤其是伺服器，然而，由於伺服器系統的複雜性和伺服器應用的多樣性，這項工作極具挑戰性。本研究提出一套開發利用機器學習自動優化伺服器應用的工具的方法，該工具能辨識伺服器應用的輸入/輸出模式（I/O pattern），並透過調整伺服器的系統和存儲裝置的設定來優化效能，而這些設定使用兩階段設定空間探索決定。該工具還備有拒絕機制，以避免對未知的模式進行不合宜的優化──因為現實世界飛速變化，不太可能涵蓋全部的應用。評估顯示，使用了該工具後，伺服器應用的效能平均提高了1.17—1.43倍。因此，本研究可以讓系統管理員毋須具備過多關於底層的系統和存儲裝置、甚至是工作負載本身的知識，便能優化伺服器應用的效能，此外，研究人員和系統開發人員更可將這套方法應用於其他平台。本研究亦針對伺服器效能優化探索了數個有潛力的機器學習技術，為未來的工作提供了寶貴的見解。

Performance tuning is critical to contemporary computer systems, especially servers. However, it is challenging due to the complexity of server systems and the diversity of server applications. This study proposes a methodology for developing a tool that leverages machine learning (ML) to automatically optimize server applications. The tool recognizes the server application’s I/O pattern and optimizes the performance by tuning the server’s system and storage configuration, which is determined with the two-stage configuration space exploration. Besides, the tool is equipped with a rejection mechanism to avoid performing undesired optimization for unknown patterns, since it is impossible to encompass all applications in a fast-changing world. The evaluation shows that after using the proposed tool, the performance of the server applications has a 1.17-1.43x speedup on average. As a result, the study can help system administrators tune workload performance without requiring much knowledge about the underlying system and storage and the workload itself. Researchers and system developers may apply the methodology to other platforms. Furthermore, this study explores different promising ML techniques for server performance tuning, providing valuable insights for future work.