Smartbuf: An Agile Memory Management for Shared-Memory Switches in Datacenters
Hamed Rezaei, Hamidreza Almasi, Balajee Vamanan · 2021
Important datacenter applications generate extremely bursty traffic patterns and demand low latency tails as well as high throughput. Datacenter networks employ shallow-buffered, shared-memory switches to cut cost and to cope up with ever-increasing link speeds. End-to-end congestion control cannot react in time to handle bursty, short flows that dominate datacenter traffic and they incur buffer overflows, which cause long latency tails and degrade throughput. Therefore, there is a need for agile, switch-local mechanisms that quickly sense congestion and provision enough buffer space dynamically to avoid costly buffer overflows. We propose Smartbuf, an online learning algorithm that accurately predicts buffer requirement of each switch port before the onset of congestion. Our key novelty lies in fingerprinting bursts based on the gradient of queue length and using this information to provision just enough buffer space. Our preliminary evaluations show that our algorithm can predict buffer demands accurately within an average error margin of 6% and achieve an improvement in the 99thpercentile latency by a factor of 8x at high loads, while providing good fairness among ports.