Towards High Performance Networked Systems

Опубликовано 21 июня 2016, 19:29

In this talk, I will discuss a few projects that my research group has recently worked on. In the first half of my talk, I will describe our research on designing protocols for achieving low latency for data center (DC) applications. I will first describe a deployment-friendly transport for private and enterprise DCs that uses adaptive control laws to dynamically prioritize short flows over long flows. I will then discuss PASE; a new transport framework that achieves near-optimal performance while being deployment-friendly. PASE employs a novel control plane for arbitration between flows. It uses priority queues available in existing switches for flow prioritization at per-packet timescales and employs TCP-style endpoint-based probing to achieve work conservation. Finally, I will discuss a buffer management framework for public clouds for achieving high performance for data center flows. In the second half, I will describe BLMon, a low overhead framework for loss differentiation in high-speed WLANs. A fundamental problem in 802.11 wireless networks is to determine the cause of packet losses. This becomes increasingly important as wireless data rates scale to Gbps. Recent and upcoming high-speed WLAN standards, such as 802.11n and 802.11ac, use frame aggregation and block acknowledgements for achieving efficient communication. We observe that noise, collisions, and hidden nodes result in distinct loss patterns within aggregate frames and aggregate frame retries. BLMon leverages this structure to achieve accurate and low overhead loss differentiation. We implement BLMon in the ath9k driver and using real testbed experiments, show that it can provide up to 5x improvement in throughput. Finally, I will discuss other research initiatives being taken in my group at LUMS