[Paper Review] Dual-structure Data Center Multicast Using Software Defined Networking

Wenzhi Cui, Chen Qian|arXiv (Cornell University)|Mar 31, 2014

Software-Defined Networks and 5G27 references15 citations

TL;DR

This paper proposes Dual-Structure Multicast (DuSM), a scalable multicast solution for data center networks using Software Defined Networking (SDN). DuSM classifies multicast groups into 'elephant' (high-volume) and 'mice' (low-volume) groups, using multiple shared trees for load balancing in elephant groups and state-free multicast (e.g., multicast-to-unicast translation) for mice groups. The approach increases multicast group support by 300% compared to IP multicast while achieving better traffic load balancing across links.

ABSTRACT

Data center applications use multicast as an effective method to reduce bandwidth cost. However, traditional multicast protocols designed for IP networks are usually bottlenecked by the limited state capacity on switches. In this paper, we propose a scalable multicast solution on fat tree networks based on the observation that data center multicast traffic has strong heterogeneity. We propose to remove the multicast management logic from switches and use the SDN controller to manage multicast groups. The proposed Dual-structure Multicast (DuSM) determines elephant and mice groups according to their traffic amounts and treats them separately. For each elephant group, the controller installs multicast state to maintain multiple shared trees and the group traffic will be balanced evenly among the trees to avoid congestion. For mice groups, the controller applies state-free mutlicast that trades bandwidth capacity for state capacity, such as multicast-to-unicast translation. Our experiments using real multicast traffic data show that the number of groups DuSM supports can be 300% of that of IP multicast. DuSM also achieves traffic balance among links.

Motivation & Objective

• To address the scalability limitations of traditional IP multicast in data center networks due to limited multicast state capacity on switches.
• To leverage traffic heterogeneity in data center multicast—where a few groups dominate bandwidth usage—by classifying groups into elephant and mice categories.
• To design a centralized, SDN-based multicast management system that optimizes state usage and traffic balancing across links.
• To reduce network congestion and improve bandwidth efficiency by using multiple shared trees for high-volume groups and state-free techniques for low-volume groups.

Proposed method

• The SDN controller collects network information and classifies multicast groups based on traffic volume, distinguishing between elephant and mice groups.
• For elephant groups, the controller installs multiple shared multicast trees (Steiner trees) to distribute traffic evenly and prevent congestion.
• For mice groups, the controller applies state-free multicast techniques such as multicast-to-unicast translation to conserve multicast state space.
• The system uses a centralized controller to compute and install forwarding rules on switches, enabling dynamic and topology-aware multicast routing.
• The design is evaluated using real multicast traces from data centers, with implementation and testing in the ns-2 simulator.
• The approach leverages the fat tree topology’s multiple parallel paths to improve load balancing and avoid bottlenecks.

Experimental results

Research questions

• RQ1 Can a dual-structure approach that differentiates between high-volume and low-volume multicast groups improve multicast scalability in data center networks?
• RQ2 How can SDN be used to centralize multicast group management and reduce reliance on switch-level multicast state?
• RQ3 To what extent can state-free multicast techniques for mice groups reduce multicast state space without degrading bandwidth efficiency?
• RQ4 Can multiple shared multicast trees effectively balance traffic and prevent congestion in high-volume (elephant) groups?
• RQ5 How does the proposed system compare to traditional IP multicast in terms of supported multicast groups and link load balance?

Key findings

• DuSM supports up to 300% more multicast groups than traditional IP multicast by reducing multicast state usage through classification and state-free techniques.
• The system achieves better traffic load balancing across network links by distributing elephant group traffic across multiple shared trees.
• State-free multicast for mice groups significantly reduces the multicast state space required on switches, enabling scalability on commodity hardware.
• The use of a centralized SDN controller allows dynamic and topology-aware multicast routing, improving overall network efficiency.
• The approach reduces controller overhead for switch updates, mitigating a potential scalability bottleneck in large-scale deployments.
• Experimental evaluation using real data center traffic traces confirms the effectiveness of DuSM in enhancing scalability and load balancing.