CoNEXT'13: Scaling IP Multicast on Datacenter Topologies

Speaker: Xiaozhou Li
Author: Michael J. Freedman 

This paper proposes a mechanism to increase the number of available multicast group in a Data Center.

Why IP multicast is difficult for Data Centers?

It has scaling difficulties at control plane and data plane.

Difficulties come from limited forwarding table size (limited memory size).

Multicast addresses cannot be aggregated by prefixes and must maintain per-group forwarding rules for all groups.

This paper proposes three techniques to increase multicast group:

1.       Partition and distribute multicast address space

2.       Enable local multicast address aggregation to further increase the number of groups in each pod

3.       Handle failures with fast rerouting and multicast tree reconstruction

Each core and aggregate switches has a portion of the partition and name space (with fix prefix)

They needed to cope with the problem of bottleneck problem at aggregate switch

Solution:

·         Reduce the number of entries in the bottleneck switch

·         Local address translation and aggregation

Compute the aggregation is NP-hard

Approach:

·         Local aggregation at the bottleneck

They also provide fault tolerance

They use SDN to manage multicast memberships.

Describe the simulation environment

Aim: support large num of IP-multicast groups in data centers

Contribution

Leverage multi-rooted topologies to scale out by dividing multicast address space across multiple switches.

Introduce local aggregation algorithm to overcome bottleneck in pods.

Proposed mechanism for fast failure and multicast tree management practical with today’s SDN

Q: we talked about the related work which previous works use bloom filters on SDN and you said they cannot support very large scale networks but you don’t have any simulation results compare with the bloom filtered design.

A: So they can increase the number of groups in one switch but we are not looking at the single switch and compare to us they have much lower increase in number of groups. And these two approaches can be combined together.   

Q: So you can combine bloom filter with your scheme?

A: maybe not bloom filter but some other previous ideas can be combined with ours.

Q: Just as a suggestion you can try bloom filter schemes to see how much performance you have.