Multicast Optimization for CLOS Fabric in Media Data Centers

Abstract

Multicast is widely deployed in data centers for point-to-multi-point communications. Multicast is increasingly being used to carry uncompressed video in media data centers with very large bandwidth requirements per flow. Multicast control protocols such as IGMP and PIM build multicast trees without trying to maximize the overall fabric capacity, leading to decreased fabric utilization and inability to service flows. In addition, existing multicast protocols are not bandwidth-aware and could cause links to over-subscribe leading to packet loss and negative user quality of experience. In this paper, we formulate offline optimization for multicast trees in clos fabric. We then design and implement two novel algorithms, iRP and LiRP, to optimize multicast tree formation and increase overall fabric multicast capacity. iRP algorithm optimizes online formation of multicast trees while addressing bandwidth requirements using SDN controller. We share analysis of TV studio repetitive traffic patterns the benefits of time series forecasting to predict multicast group membership and bring online optimization efficiency closer to offline optimization results. We then implement and test LiRP algorithm to increases iRP's fabric efficiency by implementing k-fold cross validation method to predict future multicast group memberships leading to optimized multicast tree placement. We implement iRP and LiRP algorithms using controller-based system and test both algorithms using Cisco Nexus commercially available switches. Testing results confirm that iRP Algorithm increases fabric capacity by 60% compared to PIM performance. LiRP system increases the efficiency of iRP by up to 40% through prediction of multicast group memberships with online arrival.