2014 Poster Sessions : Improving Datacenter Network Performance with Multicast

Student Name : Michael Chan
Advisor : David Cheriton
Research Areas: Computer Systems
Abstract:
Data replication is a common activity in data centers. For example, applications depend on efficient fault-tolerant data replication provided by distributed file systems. As large-scale data analytics take on an increasingly important role in businesses, one could expect more frequent fresh data ingestion into compute clusters, increasing the network overhead brought about by replication. IP multicast allows data replication in the network at line rate, but critically does not perform congestion control. We propose a straightforward enhancement to TCP -- TCP Multicast - so that multicast is effectively congestion controlled even at high line rate (10Gbps). Our results show that with multicast replication, one can scale distributed file system read/write rates linearly with offered load, and reduce read/write latency by 50% compared to existing unicast-based replication schemes. TCP Multicast is also extended to support queue-aware congestion control algorithms like Datacenter TCP. This allows throughput-oriented multicast flows to be run without degrading the performance of concurrent, short-lived latency-sensitive flows.

Bio:
Michael Chan is a computer science Ph.D. candidate working with Prof. David Cheriton at Stanford University. His research is focused on improving data center network performance with multicast. His work spans multiple areas, including the operating system network stack, network drivers, network switch architecture and distributed systems.