PH-distributed fault models for mobile communication

Abstract

In this paper we analyze the quality of wireless data transmission. We are primarily interested in the importance of the distance between sender and receiver when measuring data loss rate and the length of lossy and loss-free periods. The ultimate purpose of this type of study is to quantify the effects of mobility. We have sampled data and find that distance certainly is an important indicator but the loss rate of packets is also determined by other factors and does not always monotonically increase with the distance. We further find that while the distribution of the length of lossy periods mostly shows an exponential decay the distribution of the length of loss-free periods does not even always monotonically decrease. Both, the packet loss probability and the distribution of the length of loss-free periods can be well represented using probabilistic models. We fit simple Gilbert-Elliot models as well as phase-type distributions to the data using different fitting tools and provide loss models that can easily be used in simulation and testbed studies.

Fault-Impact Models Based on Delay and Packet Loss for IEEE 802.11g

“…Carvalho et al claim that the Gilbert model cannot capture certain characteristics of wireless errors in 802.11g networks. Wolter et al [27] observed the same. That would suggest that links experience burst errors which contradicts some previous studies [1].…”

“…In [27], we observed that there is a strong correlation between distance and average PDR. Our current results do not fully support this observation.…”

“…However, our current measurement setup in the DES testbed is quite different from the one we used in [27]. In particular, the current setup involves a large set of nodes in separate buildings under real-life operating conditions, whereas in [27] we used an isolated setup of 2 nodes. Figure 3 illustrates the spatial distribution of link qualities based on the average PER.…”

“…Several experimentation studies on wireless network behaviour have been conducted in the past [1,4,11,17,27] and evaluated in terms of standard analysis techniques such as moments and histograms. These studies provide insight into network behaviour and the traces can be used in detailed simulation models.…”

“…While in the simplest case packet loss may be described by a Bernoulli model, packet loss is often comprised of bursts of elevated loss probability, which can be modelled more closely with Gilbert-Elliot (GE) models [5][6][7]11,12,17,[27][28][29]. A Gilbert-Elliot model describes the loss process as a Markov Chain with different loss probabilities for each state.…”