Three Perspectives to Enhance the Performance of High-Speed Datacenter Networks

Speaker : Vamsi Addanki
TU Berlin
Date: 17/05/2023
Time: 3:00 pm - 4:00 pm
Location: Room 4B01

Abstract

The performance of large-scale computing systems often critically depends on high-performance communication networks. In this presentation, I will focus on two key metrics throughput and latency from three perspectives within a datacenter network: end-host, network dataplane and topology.

At the end-host, I will discuss the impact of congestion control in achieving high throughput and low latency while requiring little to no buffer in the network. I will then present PowerTCP, a power-based congestion control algorithm that leverages inband network telemetry and adapts to both the absolute state as well as the changes in the network (NSDI 2022).

At the network core, I will discuss the role of buffer management in improving datacenter performance and present Active Buffer Management (ABM), a novel buffer sharing scheme for datacenter switches.  ABM not only offers isolation but also offers stable queueing delays which as we show essentially enhances the burst absorption capabilities of the switch (SIGCOMM 2022).

At the topological level, I will show the fundamental tradeoffs across throughput, latency and buffer requirements within a subset of the design space. I will briefly present Mars, a near-optimal throughput reconfigurable datacenter topology that achieves high throughput within the given buffer capacity of the network devices (SIGMETRICS 2023).

I will conclude by discussing ongoing works and open questions.

 

Bio

Vamsi is a third year PhD student at TU Berlin under the supervision of Prof. Stefan Schmid.  He obtained MSc. from Sorbonne University Paris and did his thesis at ETH Zurich. In 2017, he started off his path towards research at Telecom Paris with Prof. Dario Rossi and Leonardo Linguaglossa hosted at Lincs. Currently, he is broadly interested in network topologies; congestion control and network flow problems; online algorithms and competitive analysis; buffer requirements and management; and related areas. His research is mainly at the intersection of networking theory and systems.