High throughput path selection for unstructured data center networks

Sharma, Abhishek

Permalink

https://hdl.handle.net/2142/45451 Copy

Description

Title

High throughput path selection for unstructured data center networks

Author(s)

Sharma, Abhishek

Issue Date

2013-08-22T16:40:33Z

Director of Research (if dissertation) or Advisor (if thesis)

Godfrey, Philip B.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Data Center
• Multipath Transmission Control Protocol (MPTCP)
• Jellyfish
• Shortest Paths
• Disjoint Paths
• Throughput
• Random Regular Graph (RRG)
• Random Permutation Matrix
• Switches
• Servers
• Ports
• Network Density

Abstract

The increase in demand and popularity of cloud and big data applications has driven the need for higher throughput data center network design. Recent work to provide topologies with much denser interconnects pose a difficult challenge for routing of traffic within a data center. Even with proposals like MPTCP that improve upon TCP, there is a still a gap between theoretical throughput and empirical throughput. Our goal is to study routing in data centers at both flow and packet-level to determine the cause of inefficiency and study methods of path selection that may help bridge the gap between optimal throughput and packet level throughput. The difference in throughput can be attributed to inefficiency due to path selection and inefficiency due to protocol overhead; we quantify the contribution of each. Focusing on path selection, our experiments show that k-disjoint paths provide much better throughput in most topologies than the previously used k-shortest paths. We also show that one can positively impact network throughput by varying the number of paths according to network density.

Graduation Semester

2013-08

Permalink

http://hdl.handle.net/2142/45451

Copyright and License Information

Copyright 2013 Abhishek Sharma. All rights reserved

Owning Collections