無資料遺失之重新啟動架構在網際網路服務上之設計與實作

Abstract

近年來網際網路上的各種服務愈來愈被廣泛的應用來解決問題或尋找資料。這也使得這些網際網路服務更容易因為長期的運作，或是同時服務大量的客戶，而發生突發性的暫時錯誤。因此，我們提出了一個無資料遺失的重新啟動架構，使得網際網路服務能夠永續的運作。 以往已有一些無資料遺失的重新啟動方法，例如檢查點的建立技術，但是這會造成不小的網際網路服務運作效能減低，而且也並不完全適用於解決突發性的暫時錯誤。另外，FT-TCP也一個是使網際網路服務能不中斷的方法，但是再重新啟動的時候可能會耗費較多的時間。而我們的架構能夠達到一樣的效果但是又有使得整體效能降低不多，而且重新啟動的時候也不會耗費太多的時間。 我們的架構包含了兩個部分。首先，需要修改網際網路服務使他的運作狀態能被存放在安全的地方，這是為了避免重新啟動之後的資料流失。另外，我們還提供了一些系統核心上的支援，使得網際網路服務所用到的輸入輸出頻道能被保留並且自動的完成整個重新啟動的流程。最後，我們將這個架構套用在一個熱門的網頁伺服器上，並且驗證了我們的想法。

Internet services are more and more widely used for resolving problems or finding information recently. Thus, Internet services have to run for a long time and must serve a huge number of clients concurrently. This causes them easy to suffer from the transient fault and software aging problem, which will cause serious damage when them occurs at some critical Internet services such as commercial web sites. In order to keep Internet service permanently running, we propose a zero-loss restart framework to resolve the transient fault or aging problem of Internet services. Our framework consists of three parts, the service state abstraction, the kernel I/O channel keeping kernel support, the automatic restarting mechanism kernel support. In addition, we deploy our framework on a popular tiny web server, thttpd. The former zero-loss restart mechanism such as checkpoint and restart has the shortcoming of high runtime overhead and can not totally resolve the transient fault. Another popular zero-loss restart mechanism is the FT-TCP which also has the shortcoming of high recovery time. As result of experimental, our framework only costs lower than 6.8% runtime overhead and lower than 8.5% recovery overhead to perform the zero-loss restart.