Instrumentation & Measurement Magazine 25-6 - 25

Network Delay Measurement with
Machine Learning: From Lab to
Real-World Deployment
Shady A. Mohammed, Shervin Shirmohammadi, and Alaa Eddin Alchalabi
A
rtificial Intelligence (AI) continues to impact all facets
of technology including Instrumentation and
Measurement (I&M) with much effort spent on developing
I&M systems assisted by machine learning (ML),
especially deep learning [1]. While these ML-assisted I&M
systems show promising results in a lab environment, there is
always the question of how well they will perform in the real
world. In fact, concerns about the real-world performance of
ML is not exclusive to I&M but an inherent property of ML in
general, because ML is data driven and its performance will
change if the data distribution changes in the real world. In
this article, we present a case study of developing in the lab an
ML-assisted I&M system, specifically a network delay predictor,
and deploying it in the real world, achieving 93% accuracy.
Network Delay Prediction versus Direct
Measurement
Although the delay between two nodes on a computer network
can be directly measured, there are applications in which
direct measurement is not practical. In a large-scale N-node
network, the computational overhead, O(N2
), and the network
traffic overhead generated by probing, make direct measurement
between every pair of nodes impractical if N is large. For
instance, one objective in cloud or edge computing is the efficient
allocation of incoming users to individual cloud or edge
servers such that not only the user's quality of experience is
maximized but also the provider's costs are minimized [2]. In
cloud gaming, for example, an incoming player should be assigned
to an edge server that gives the player the lowest delay
[3], but the massive number of players makes it impractical
for direct measurement of network delay between each player
and all servers. Other examples are content distribution networks
and peer-to-peer networks.
For these cases, predicting the network delay with good
accuracy and without actual measurement is an attractive solution.
In [4], we presented a deep-learning-assisted network
delay predictor, which we call Deme (Delay Measurement Estimator),
that outperforms existing analytical solutions in both
accuracy and response time. While the results were promising,
it remained to be seen how well Deme performs in the real
world, which is the subject of this article. But before we can
present that, a basic understanding of the inner workings of
Deme is necessary, and this is described next.
In the Lab: Design and Implementation
of Deme
The following is a summary of Deme's architecture and
design. For more details, interested readers are referred
to [4]. As can be seen from Fig. 1, Deme takes only two inputs,
source and destination nodes' IP addresses, which
are then used to extract features such as VPN, DNS, ASN,
and geo-coordinates via a service such as KeyCDN. Afterwards,
word embedding is performed to convert strings to
numbers, and geo-distance is calculated as an additional
Fig. 1. Deme's overall architecture. ©IEEE, reused with permission from [4].
September 2022
IEEE Instrumentation & Measurement Magazine
1094-6969/22/$25.00©2022IEEE
25
Instrumentation & Measurement Magazine 25-6

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-6
