Elasticity Detection: A Building Block for Delay-Sensitive Congestion Control
Name
1802.08730.pdf
Description
Accepted version
Size
1.56 MB
Format
Adobe PDF
Checksum (MD5)
320c569b83a56647dd7f288fada4cd24
Author(s)
Goyal, Prateesh
Narayan, Akshay
Cangialosi, Frank
Raghavan, Deepti
Narayana, Srinivas
Alizadeh, Mohammad
Balakrishnan, Hari
Date Issued
July 16, 2018
Publisher
ACM
Citation
Goyal, Prateesh, Narayan, Akshay, Cangialosi, Frank, Raghavan, Deepti, Narayana, Srinivas et al. 2018. "Elasticity Detection: A Building Block for Delay-Sensitive Congestion Control."
Version
Author's final manuscript
Abstract
This paper develops a technique to detect whether the cross traffic competing with a flow is elastic or not, and shows how to use the elasticity detector to improve congestion control. If the cross traffic is elastic, i.e., made up of buffer-filling flows like Cubic or Reno, then one should use a scheme that competes well with such traffic. Such a scheme will not be able to control delays because the cross traffic will not cooperate. If, however, cross traffic is inelastic, then one can use a suitable delay-sensitive congestion control algorithm, which can control delays, but which would have obtained dismal throughput when run concurrently with a buffer-filling algorithm.
We use the elasticity detector to demonstrate a congestion control framework that always achieves high utilization, but which can also achieve low delays when cross traffic permits it. The technique uses an asymmetric sinusoidal pulse patternand estimates elasticity by computing the frequency response(FFT) of the cross traffic estimate; we have measured its accuracy to be over 90%. We have developed Nimbus, a protocol that explicitly switches between TCP-competitive and delay-sensitive modes using the elasticity detector. Our results on emulated and real-world paths show that Nimbus achieves throughput comparable to or better than Cubic always, but with delays that are much lower when cross traffic is inelastic. Unlike BBR, Nimbus is fair to Cubic, and has significantly lower delay in all cases; for example, on real-world paths, Nimbus has 11% lower throughput but at 40-50 ms lower packet delay.
MIT Department
Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1145/3232755.3232772
