BurstRadar
Raj Kumar Joshi, Ting Qu, Mun Choon Chan, Ben Leong, Boon Thau Loo · 2018
Microbursts can degrade application performance in datacenters by causing increased latency, jitter and packet loss. The detection of microbursts and identification of the contributing flows is the first step towards mitigating this problem. Unfortunately, microbursts are unpredictable and typically last for 10's or 100's of μs and the high line rates (> 10 Gbps) in modern datacenter networks further exacerbate the problem. In this paper, we show that modern programmable switching ASICs have made it practical to detect and characterize microbursts at high line rates. Our system, called BurstRadar, operates in the dataplane and monitors microbursts by capturing the telemetry information for only the packets involved in microbursts. We have implemented a prototype of BurstRadar on a Barefoot Tofino switch using the P4 programming language. Our evaluation on a multi-gigabit testbed using microburst traffic distributions from Facebook's production network shows that BurstRadar incurs 10 times less data collection and processing overhead than existing solutions. Furthermore, BurstRadar can handle simultaneous microburst traffic on multiple egress ports while consuming very few resources in the switching ASIC.