We all agree that voice wireless service providers have based their measurements on general indicators like cell coverage and capacity in order to assess their network performance. Measurements such as drop call rate, block call rate or voice quality are used to quantify such performance.
These voice performance indicators are based on specific characteristics of Circuit Switched (CS) traffic, such as dedicated channels for each user and no error correction in case of frame loss. When moving to the area of data services, traffic sources present very different and variable profiles. Users can share the resources of same data channel, the rate of the assigned channel may be changed according to user conditions, and retransmissions can be performed in case of transmission errors. This variability in the traffic distribution makes it difficult to define common measurements which can give, at the same time, the performance of the network and the user perception of each service. A clear example can be seen in traditional network monitoring, which averages the performance of different traffic profiles (Web download, streaming, e-mail, etc.), by aggregating reported events over certain period. The typical reporting periods of hours, days or weeks, average not only over the time, but also over different services, file sizes, different burstyness, etc. The result is that the indicators obtained are so mixed that they lose the correlation with individual user experience.
As we all know Legacy cellular networks do not implement a full end-to-end QoS approach. They operate just as an additional node integrated in the whole Internet, where the access point to the wireless users is a router (the GGSN) which separates the Internet domain from the operators’ domain. Inside the operators’ domain, a series of mechanisms aimed to provide QoS assurance may be implemented according to the 3GPP and 3GPP2 specifications, but part of the performance will depend on the implementation of the external network and the protocols that are used. The interface to the external network from the layering point of view is the IP level, which will be affected by routing policies and capacities, but the control of the transport and application layers will depend very much on the configuration of the end peers, typically client and server, and the version of the protocol used.
This method follows a bottom-up approach, starting from the lower levels of the layer architecture and considering a cumulative degradation of the performance based on the effects of the different layers and their interactions. The ideal throughput provided by layer one (physical layer) is considered initially as the starting point, and then the performance degradation introduced by each of the upper layers in the protocol stack is estimated. The way in which the performance of a certain service is degraded, compared to others, will depend also on the kind of service and the factors towards it is more sensitive: throughput, delay, response time, packet loss, etc.
The illustration below presents, as an example, the different protocol stack layers in GERAN networks.
From a service performance point of view, the set of factors that produce a degradation of the link level throughput can be grouped into data link level and upper layer effects.
Data link effects are those factors affecting the performance that depend on the network conditions, such as interference, coverage, radio resource sharing, etc. The resulting performance after these effects is the data link throughput, which refers to the payload (throughput) offered by the Radio Access Network (RAN) to the transport and application layers. Data link throughput and the network latency are the basic elements that can be calculated for different networks and used as a starting point to estimate the performance of each service. Both measurements depend only on the network itself.
The upper layer effects are those factors which depend on the transport and application protocols. They contribute to the degradation of the user service performance in the same way, independently on the network that is used for the transmission. What determines the proportional impact of these effects on the final service performance are both the throughput and the latency at the data link level. Normally networks with higher latencies will suffer much more from these degradations than networks with less latency, even if the data link layer throughputs are quite similar. Based on these service performance characterization I include here an illustration showing the main sources of performance degradation in (E)GPRS networks.
