ESIR: A Deployment System for Large-Scale Server Cluster

Deploying a large-scale server cluster is troublesome job for administrators. Aiming at efficiency, scalability, independence and reliability, we design and implement a deployment system for large-scale server cluster called ESIR. ESIR is an image-based deployment system, and it leverages file mode method to make images. Dynamic module loading technique has been proposed to implement hardware-independent deployment. We introduce BitTorrent (BT) to ESIR to achieve efficient and reliable system image transfer. To avoid the HAVE messages storm in high bandwidth cluster environment, we proposed adaptive HAVE messages mechanism which reduces about 90% HAVE messages of BT per second without delay of exchanging newly downloaded piece information. The measurement result on image making method shows system image size made by file mode is much smaller than that by partition mode, which significantly reduces the deployment time. Measurement comparison on multicast and improved BT shows that the later promises better reliability and platform independent. The measurement on performance indicates that improved BT has higher image distribution performance than reliable multicast, it deploys 596MB system image to 32 servers in no more than 260 seconds.