Packet-based load balancing in data center networks

Yagiz Kaymak, Roberto Rojas-Cessa

Research output: Contribution to journalArticlepeer-review

1 Scopus citations

Abstract

In this paper, we evaluate the performance of packet-based load balancing in data center networks (DCNs). Throughput and flow completion time are some of the main the metrics considered to evaluate the performance of the transport of flows over the presence of long flows in a DCN. Load balancing in a DCN may improve those metrics but it may also generate out-of-order packet forwarding. Therefore, we investigate the impact of out-of-order packet delivery on the throughput and flow completion time of long and short flows, respectively, in a DCN. We focus on per-packet load balancing. Our simulations show the presence of out-of-order packet delivery in a DCN using this load balancing approach. Simulation results also reveal that packet-based load balancing may yield smaller average flow completion time for short flows and larger average throughput for long flows than the single-path transport model used by Transmission Control Protocol (TCP), which prevents the presence of out-of-order packet delivery. Queueing diversity in the multipath structure of DCNs promotes susceptibility of out-of-order delivery. As the delay difference between alternative paths decreases, the occurrence of out-of-order packet delivery in packet-based load balancing also decreases. Therefore, under the studied scenarios, the benefits of the packet-based load balancing seem to outweigh the out-of-order problem.

Original languageEnglish (US)
Pages (from-to)1-18
Number of pages18
JournalJournal of Cyber Security and Mobility
Volume5
Issue number1
DOIs
StatePublished - Jan 1 2016

All Science Journal Classification (ASJC) codes

  • Hardware and Architecture
  • Computer Networks and Communications

Keywords

  • Data center network
  • Flow completion time
  • Load balancing
  • Multipath forwarding
  • TCP

Fingerprint

Dive into the research topics of 'Packet-based load balancing in data center networks'. Together they form a unique fingerprint.

Cite this