Low-Cost Datacenter Load Balancing With Multipath Transport and Top-of-Rack Switches

Load balancing in datacenter networks (DCNs) is an important and challenging task for datacenter managers. A number of sophisticated technologies have been proposed to improve load balancing performance in a complicated circumstance, i.e., with various traffic characteristics. Many approaches need a...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on parallel and distributed systems 2020-10, Vol.31 (10), p.2232-2247
Main Authors: Dong, Enhuan, Fu, Xiaoming, Xu, Mingwei, Yang, Yuan
Format: Article
Language:English
Subjects:
Data transmission
Datacenter networks
Load balancing
Local flow
Low cost
multipath TCP
network communications
network protocols
Simulation
Switches
Traffic congestion
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Load balancing in datacenter networks (DCNs) is an important and challenging task for datacenter managers. A number of sophisticated technologies have been proposed to improve load balancing performance in a complicated circumstance, i.e., with various traffic characteristics. Many approaches need a high cost to implement, such as changing switch hardware. The efficiency problem has not been well addressed. MPTCP was proposed as a low-cost approach to improve data transmission in DCNs, which uses subflows to balance workloads across multiple paths. However, current MPTCP is not satisfying, especially when there are rack-local flows or many-to-one short flows. In this article, we propose DCMPTCP to improve the efficacy of MPTCP. We gradually develop three mechanisms. First, DCMPTCP identifies rack-local traffic and eliminates unnecessary subflows to reduce the overhead. Second, DCMPTCP estimates flow length and establishes subflows in a smarter way. Third, DCMPTCP strengthens explicit congestion notification to improve the congestion control performance on inter-rack many-to-one short flows. We have implemented DCMPTCP in both the Linux kernel and ns-3 simulator. Our comprehensive testbed experiments and simulations show that DCMPTCP outperforms MPTCP in both 1 Gbps testbed, and 10 Gbps large-scale simulation network.
ISSN:1045-9219
1558-2183
DOI:10.1109/TPDS.2020.2989441