Narino Mendoza, Juan Pablo
[UCL]
Donnet, Benoit
[UCL]
Dupont, Pierre
[UCL]
Using quality-of-service (QoS) metrics for Internet traf-
fic is expected to improve greatly the performance of many
network enabled applications, such as Voice-over-IP(VoIP)
and video conferencing. However, it is not possible to con-
stantly measure path performance metrics (PPMs) such as
delay and throughput without interfering with the network.
In this work, we focus on PPMs measurement scalabil-
ity by considering machine learning techniques to estimate
predictive models from past PPMs observations. Using real
data collected from PlanetLab, we provide a comparison be-
tween three different predictors: AR(MA) models, Kalman
filters and support vector machines (SVMs). Some predic-
tors use delay and throughput jointly to take advantage of
the possible relationship between PPMs, while other pre-
dictors consider PPMs individually. Our current results
illustrate that the best performing model is an individual
SVM specific to each time series. Overall, delay can be pre-
dicted with very good accuracy while accurate forecasting
of throughput remains an open problem.
Bibliographic reference |
Narino Mendoza, Juan Pablo ; Donnet, Benoit ; Dupont, Pierre. A comparative study of path performance metrics predictors.Advances in learning for networking, ACM workshop in conjunction with SIGMETRICS/Performance (Seattle, USA, du 15/06/2009 au 15/06/2009). |
Permanent URL |
http://hdl.handle.net/2078.1/79323 |