FlowDyn: Towards a Dynamic Flowlet Gap Detection using Programmable Data Planes

TL;DR

FlowDyn introduces a dynamic flowlet gap detection method using programmable data planes to improve load-balancing efficiency in data center networks, reducing flow completion times significantly across different load levels.

Contribution

It presents a novel adaptive flowlet gap mechanism that leverages active probes and telemetry for better load-balancing in data center networks.

Findings

Achieves 3.19x smaller flow completion time at 10% load

Reduces flow completion time by 1.16x at 90% load

Uses programmable data planes for real-time path latency tracking

Abstract

Data center networks offer multiple disjoint paths between Top-of-Rack (ToR) switches to connect server racks providing large bisection bandwidth. An effective load-balancing mechanism is required in order to fully utilize the available capacity of the multiple paths. While packet-based load-balancing can achieve high utilization, it suffers from reordering. Flow-based load-balancing such as equal-cost multipath routing (ECMP) spreads traffic uniformly across multiple paths leading to frequent hash collisions and suboptimal performance. Finally, flowlet based load-balancing such as CONGA or HULA splits flows into smaller units, which are sent on different paths. Most flowlet based load-balancing schemes depend on a proper static setting of the flowlet gap, which decides when new flowlets are detected. While a too small gap may lead to reordering, a too large gap results in missed load-balancing opportunities. In this paper, we propose FlowDyn, which dynamically adapts the flowlet gap to increase the efficiency of the load-balancing schemes while avoiding the reordering problem. Using programmable data planes, FlowDyn uses active probes together with telemetry information to track path latency between different ToR switches. FlowDyn calculates dynamically a suitable flowlet gap that can be used for flowlet based load-balancing mechanism. We evaluate FlowDyn extensively in simulation, showing that it achieves 3.19 times smaller flow completion time at 10% load and 1.16x at 90% load.