A scalable and efficient multicast system for data center

Abstract

It is a common case in many cloud computing applications to transfer massive amount of data from a source to multiple destinations, which is usually be implemented with IP multicast. However, IP multicast is not supported in dater center, so to implement one-to-many communication in data center is an important issue. This paper addresses this problem by proposing JU, a new multicast system, which includes: 1) An application method with the topology knowledge to build and maintain multicast tree; 2) A scalable approach to forward multicast packets which is not limited by hardware; 3) A new TCP-friendly multicast transport protocol which makes the receivers get the packets as fast as they can. With simulation, we show that the multicast completion time of our approach is shorter than other technologies'. If IP multicast is deployed in data center without the limit, JU's completing time is only 50% of IP multicast. We also show that our transport protocol is fair with TCP.