Abstract
Broadcast television traditionally employs a unidirectional
transmission path to deliver low latency, high-quality
media to viewers. To expand their viewing choices, audiences
now demand internet OTT (Over The Top) streamed media with
the same quality of experience they have become accustomed to
with traditional broadcasting. Media streaming over the internet
employs elephant flow characteristics and suffers long delays due
to the inherent and variable latency of TCP/IP.
Early detection of media streams (elephant flows) as they enter
the network allows the controller in a software-defined network to
re-route the elephant flows so that the probability of congestion is
reduced and the latency-sensitive mice flows can be given priority.
This paper proposes to perform rapid elephant flow detection,
and hence media flow detection, on IP networks within 200ms
using a data-driven temporal sequence prediction model, reducing
the existing detection time by half.
We propose a two-stage machine learning method that encodes
the inherent and non-linear temporal data and volume characteristics
of the sequential network packets using an ensemble of
Long Short-Term Memory (LSTM) layers, followed by a Mixture
Density Network (MDN) to model uncertainty, thus determining
when an elephant flow (media stream) is being sent within 200ms
of the flow starting. We demonstrate that on two standard
datasets, we can rapidly identify elephant flows and signal them
to the controller within 200ms, improving the current count-minsketch
method that requires more than 450ms of data to achieve
comparable results.