2011 Poster Sessions : HULL: A High Bandwidth, Ultra-Low Latency Data Center Fabric Architechture

Student Name : Mohammad Alizadeh, Abdul Kabbani
Advisor : Balaji Prabhakar
Research Areas: Information Systems
Traditional measures of network goodness---goodput, quality of service, fairness---are expressed in terms of bandwidth. Network latency has rarely been a concern, partly because it is not easy to guarantee. Recently, however, there has been a shift in focus to latency as a primary metric for mainstream applications.

We present techniques to balance two seemingly contradictory goals: near baseline fabric latency and high bandwidth utilization. We explore the relative benefits of four techniques with increasing deployment complexity:(i) TCP with ECN, (ii) ECN with adaptive congestion response, (iii) pacing for long flows, and (iv) Phantom Queues (PQs) delivering congestion signals based on transmission rate rather than queue occupancy. Our implementation results show significant benefit from adaptive ECN response. However, packet pacing may be required to counter CPU-overhead-related optimizations such as interrupt coalescing and large send offloading prevalent in high speed servers. Indeed, pacing can reduce latency by a factor of five while fully utilizing links. PQs can further reduce latency by 50% at the cost of some throughput.

Abdul Kabbani is a 5th year Ph.D. student working on Data Center congestion control and packet transport mechanisms for minimizing latency.

Mohammad Alizadeh is a 5th year PhD candidate in the Electrical Engineering department, working with Balaji Prabhakar. His research is primarily focused on designing high performance packet transport mechanisms for data center networks. Mohammad completed his Bachelors degree at Sharif University of Technology. He is a recipient of the Numerical Technologies Fellowship, and the Caroline and Fabian Pease Stanford Graduate Fellowship.