학술논문

DeFlow: Differential flowlet switching for load balancing in datacenter networks
Document Type
Article
Source
In Computer Networks February 2025 257
Subject
Language
ISSN
1389-1286
Abstract
Multipath enhances the reliability and bandwidth of datacenter networks, but it also necessitates effective load balancing. The dynamic nature of traffic and diverse flow characteristics present significant challenges in achieving optimal load distribution. Existing schemes either result in mediocre performance or rely on hard-acquired global information, leading to poor FCT or high implementation costs. This paper introduces DeFlow, a simple yet efficient flowlet-based load balancing scheme implementable on programmable switches with a low cost. DeFlow distinguishes flowlets of large and small flows based on packet size and interval. It employs distinct scheduling strategies for throughout-sensitive large flows and latency-sensitive small flows under congestion by prioritizing the performance of small flows. Extensive experiments on NS-3 demonstrate that DeFlow consistently outperforms competing schemes across various topologies and workloads, improving the FCT of small flows and throughput of large flows simultaneously.