低負荷虛擬機器內部通訊

Abstract

複雜的應用與伺服器系統一般會利用multi-tier搭配machine redundancy來達到load balance與提升系統的reliability。這不但會造成硬體成本的提高，系統管理的困難，也會使得CPU 使用率大幅降低。除此之外，也因各系統均需要相互通訊，而使得內部與外部的通訊大量提升。虛擬機器技術提供了一個方式來降低機器數量與提升CPU 使用率。此外，它也提供了降低內部通訊負載量的機會。 一般而言，這些內部的通訊還是利用傳統的網路通訊協定來達成。然而，有許多虛擬機器實際上是處於同一個實體機器上。根據這一點，我們可以利用簡易的通訊協定來提升通訊的速度。 在此論文中，我們利用同一個實體機器的特性，設計了fast Inter Virtual-machine Communication (fast-IVC)的機制。該機制會在建立TCP/IP 連線時，主動判斷收發端是否處於同一台實體機器上。若是，則改用較簡單的通訊協定來取代原本TCP/IP通訊協定。根據實驗，利用fast-IVC傳遞資料可以提升50%到150％的效能。

Complex applications or Internet services are usually constructed by multiple tiers and machine to improve reliability. However, it not only requires high hardware cost and expensive management overhead but also incurs both inter and intra machine communication overhead. Moreover, the CPU utilization of each machine is usually low since many services are I/O-bound and each machine performs a less-complicated function in the multiple-tier architecture. The virtual machine technology was proposed for consolidating the machines and improving the CPU utilization. In addition, it also provides an opportunity to reduce the overhead of the inter-machine communication. Generally, machines communicate with each other via network protocols. However, virtual machines reside on the same physical host, making it possible to improve the performance of the communication among them. In this thesis, we propose a mechanism called fast Inter Virtual-machine Communication (fast-IVC), which transparently turns the TCP/IP based network communication between two virtual machines on the same physical host into shared memory based communication. Fast-IVC will detect whether the communication endpoints reside on the same physical machine. If they are, the complex TCP/IP stack will be skipped transparently and the communication performance will be improved. We implement the fast-IVC on Xen. According to the performance results, the performance improvement ranges from 50% to 150%.