Abstract
Mobile data services are penetrating mobile markets rapidly. The mobile industry relies heavily on data service to replace the traditional voice services with the evolution of the wireless technology and market. A reliable packet service network is critical to the mobile operators to maintain their core competence in data service market. Furthermore, mobile operators need to develop effective operational models to manage the varying mix of voice, data and video traffic on a single network. Application of statistical models could prove to be an effective approach. This paper first introduces the architecture of Universal Mobile Telecommunications System (UMTS) packet switched (PS) network and then applies multivariate statistical analysis to Key Performance Indicators (KPI) monitored from network entities in UMTS PS network to guide the long term capacity planning for the network. The approach proposed in this paper could be helpful to mobile operators in operating and maintaining their 3G packet switched networks for the long run.
1. Introduction
As mobile data usage exploded in the late 2000s, operators faced a "more with less" dilemma. Traditional network Operation and Maintenance (O&M) followed a reactive cycle: a failure occurs, a technician is dispatched, and the problem is fixed. This approach is insufficient for modern packet-switched networks where performance issues are often caused by complex traffic mixes rather than simple hardware failures.
The authors argue that simply adding bandwidth is an expensive "band-aid." Instead, operators need a way to forecast potential bottlenecks by understanding the multivariate relationships between different KPIs (latency, throughput, packet loss) and various service types (Web, VoIP, FTP, Streaming).
2. Methodology and Data
The researchers move beyond looking at single metrics in isolation, instead employing a suite of multivariate statistical tools:
- Correlation Analysis (Pearson): To measure the strength of relationships between QoS parameters and specific service types.
- Cluster Analysis & Multidimensional Scaling (MDS): Used to group different time intervals based on performance similarities, allowing for the visualization of "normal" vs. "problematic" network states in a 2D space.
- Correspondence Analysis: A graphical procedure used to associate specific time intervals with specific types of failures.
- Factor Analysis: To reduce a large set of observable variables into a few underlying "factors" that explain network behavior.
2.1 Experimental Setup
The study utilizes a simulated trial environment based on real-world data from an anonymous "Mobile Operator A." A service/call generator tool stochastically delivered this traffic into a testbed consisting of an SGSN (Serving GPRS Support Node) and a GGSN (Gateway GPRS Support Node). The monitored KPIs included GGSN utilization, packet loss on the Gn and Gi interfaces, latency, and throughput across 20 continuous sample periods.
3. Analysis and Key Findings
3.1 Service Impact and Correlations
One of the most significant findings is the statistical confirmation that "time-continuous" services (streaming video, FTP, and audio) have a near-perfect positive correlation with network latency and congestion. In contrast, "time-discrete" services like E-mail have a very low correlation (~0.17) with QoS degradation because their traffic is bursty rather than sustained. Video streaming showed the highest correlation with overall QoS degradation (~0.89).
3.2 Visualizing Network Health
By using MDS and Cluster Analysis, the authors could visualize 20 hours of complex data as distinct groups. This allows an administrator to see at a glance if the current hour is drifting toward a known "problem cluster" before a failure actually occurs. The MDS results were highly consistent with Cluster Analysis, with the first two eigenvalues explaining 82.15% of the variance, proving that a 2D visualization is sufficient for monitoring these complex networks.
3.3 Interface-Specific Diagnostics
Correspondence analysis successfully mapped specific hours of poor performance to specific interfaces. For example, it revealed that while Hour 9 and Hour 11 both suffered from packet loss, the root causes were different (Gn interface vs. Gi interface), requiring different engineering responses.
4. Conclusion and Future Work
This multivariate approach provides a robust framework for proactive network management. While the current study focused on UMTS (3G) in a controlled trial environment, the methodology is highly relevant for the transition to 4G/LTE and Evolved Packet Core (EPC) networks. Future work should explore automated policy shaping, such as dynamically adjusting Access Control Lists (ACLs) or prioritizing high-tier users during peak streaming hours identified through these statistical models.
Summary updated