以Kubernetes使用遠端虛擬記憶體與排程

遠程交換是一種能夠使用其他機器記憶體作為交換空間的一種技術，和傳統使用硬碟的交換空間相比，遠程交換有顯著的效能優勢，主要是得益於近年來高速網路和遠端直接記憶體存取的發展。目前這種方法目前僅實作於兩台機器上，在叢集上仍缺少一個有效的管理機制。為了能夠在叢集上使用遠程交換，在本篇論文中提出了一個管理系統，能夠自動在叢集內自動部署遠交換，並自動回收沒有使用到的遠程記憶體，且與能夠與當前熱門的叢集管理軟體Kubernetes一起使用，讓使用者能夠透過Kubernetes來使用叢集內的遠程交換，以及提供了一個監視介面能夠讓使用者監視當前系統內遠程交換的效能狀況。本文討論了兩種使用情境，第一種是使用遠程交換來解決大型記憶體需求的工作，當機器內的記憶體不足已完成工作時，可以使用遠程交換來滿足記憶體需求。第二種是使用遠程交換來降低等待時間，當一個工作進入叢集時，叢集內沒有一台機器有足夠的記憶體可以滿足此工作的需求時，可以使用遠程交換系統，借用別台機器閒置的記憶體來立刻執行工作。

關鍵字

遠端交換空間； RDMA ； Kubernetes

並列摘要

Remote Swapping is a technology that can use the memory in another machine as the swap space for virtual memory. Compared with a traditional swap space using hard drives, remote swapping provides significant performance improvement, thanks to the evolution of high speed network and remote direct memory access technologies in recent years. Until now, remote swapping has been set up manually between machines by the system administrators without an effective management mechanism. In this thesis, to manage the configurations and automate the establishment of the remote swapping in the cluster, we propose a remote swapping management system, which can automatically deploy the remote swapping and reclaim the unused remote memory. It works with Kubernetes, a popular cluster manage system and enables the users to take advantage of remote swapping through Kubernetes with a performance monitoring interface for diagnosing performance issues. We have carried out two case studies to demonstrate the effectiveness of this proposed work. The first case study uses remote swap to support applications with huge memory demands, allowing machines with insufficient physical memories to execute such applications effectively with remote swapping. The second case study illustrates the benefits of the proposed management scheme by reducing the wait time for the system to allocate large-memory machines, when memory-hungry applications can be executed immediately with remote swapping.