TL;DR
This paper introduces a crowdsensing-based analytical framework to pinpoint the root causes of degraded web browsing experience on mobile networks. By analyzing data from millions of smartphone users, it identifies key factors like network coverage, ISP performance, and DNS delays, offering a more user-centric approach to network optimization than traditional methods.
What Problem Does This Paper Address?
Mobile network operators traditionally focus on network-centric metrics to ensure service quality. However, with the rise of Over-The-Top (OTT) services like web browsing, this approach falls short. The user's actual experience (Quality of Experience - QoE) is influenced by a complex interplay of factors beyond just network infrastructure, including the user's device, the content provider (ISP), and even the time of day. This disconnect between network quality and user perception leads to dissatisfaction and customer churn. The paper aims to bridge this gap by developing a method to analyze and diagnose the causes of degraded web browsing QoE from the end-user's perspective, moving beyond traditional network-centric monitoring.
Methods
The core of the paper is a novel analytical framework for identifying and quantifying the impact of various factors on web browsing service perception. This framework utilizes a crowdsensing approach, collecting data directly from users' smartphones. Key methodological components include:
- Mobile Crowdsensing (MCS) Data Acquisition: A custom agent embedded in a smartphone application collects real-time data on web browsing behavior and wireless environmental parameters.
- Key Quality Indicator (KQI) Definition: Specific metrics for web browsing are defined, such as First Packet Delay ($D_k$) and Page Delay ($D_p$), which are further broken down into components like DNS resolution delay ($D_{dns}$), TCP connection setup delay ($D_{tcp}$), and GET request delay ($D_{get}$).
- Disqualification Thresholds: Statistical methods (like Q3 + 1.5IQR) are used to define thresholds for KQIs, flagging instances where performance falls below acceptable levels. These thresholds are tailored for HTTP and HTTPS websites due to their differing complexities.
- Root Cause Analysis: The framework analyzes disqualified samples across several dimensions: ISP, IP of CDN nodes, segmented delays (DNS, connection, request, response), temporal (time of day), coverage (RSRP, RSRQ), terminal (phone model), and user.
- Degradation Analysis: For each dimension, a "Ratio of Disqualified Samples" (RDS) is calculated. The framework then identifies "Disqualified" entities (e.g., ISPs, IPs, TACs) whose RDS exceeds a predefined threshold.
- Factor Significance Calculation: A normalized "Factor of Significance" (FoS) is computed to rank the impact of each dimension on overall service degradation. This involves re-labeling disqualified samples based on each factor and measuring the change in overall degradation.
- Correlation Analysis: Pearson Correlation Coefficient (PCC) and Maximum Information Coefficient (MIC) are used to quantify relationships between KQIs and factors like service intensity.
Datasets
The study utilizes a large-scale crowdsensed dataset collected from the LTE network in Gansu province, China, during August 2017. This dataset comprises:
- Volume: 711,714 samples.
- Users: Contributed by 120,304 smartphone users.
- Geographic Scope: Data collected across 14 cities in Gansu.
- Target Websites: The nine most popular websites in China (Sina, Baidu, Sohu, Taobao, People, iFeng, Tencent, Weibo, and Netease).
This dataset is a newly collected benchmark for analyzing OTT web browsing perception in a real-world mobile network environment.
Key Insights and Ideas
- User-Centricity is Paramount: The traditional network-centric view is insufficient. Understanding user perception requires data directly from their devices.
- Web Browsing is Complex: OTT web browsing is significantly more vulnerable to network variations than conventional web browsing due to its reliance on the entire mobile network chain (UE, RAN, CN, ISP).
- Coverage is King (for Wireless): Poor signal strength (RSRP) and quality (RSRQ) directly and significantly degrade web browsing KQIs like first packet delay and page delay.
- Time Matters: Service intensity, which correlates with network load, has a significant impact on web browsing performance. Peak hours often see degraded performance.
- ISP and DNS are Critical Bottlenecks: The study highlights that ISP performance (including CDN deployment) and DNS server responsiveness are major contributors to web browsing degradation, especially in developing regions.
- Terminal Fragmentation: Differences in smartphone hardware and software configurations can lead to varying service perceptions, though this factor's impact is generally less than network-related issues.
Key Results
- The framework successfully identified Sina and Tecent as "Disqualified ISPs" in the Gansu case study, with their service RDS exceeding the overall service RDS.
- Analysis revealed that CDN nodes for disqualified ISPs were located far from Gansu, indicating a lack of local CDN deployment as a major issue for that region.
- DNS delay was identified as a significant contributor to segmented delay issues, with a specific IP (202.100.64.68) flagged for potential improvement.
- Temporal analysis showed that the period from 18:00-24:00 had higher RDS values, marking it as a "Disqualified Time Period."
- Coverage (TAC areas) and Terminal (phone models) were also identified as significant factors, with 41 TAC areas and 74 phone models flagged.
- The normalized factors of significance indicated that Coverage, ISP, and DNS were the top three impact factors for overall service RDS in the Gansu province, accounting for a substantial portion of the degradation.
Limitations and Future Work
- Data Acquisition Agent Deployment: A challenge is placing the data acquisition agent onto users' smartphones, requiring careful consideration of privacy and user experience.
- Limited Network Data: Data collected at the OS level has limitations; air interface signaling data, crucial for network optimization, is not acquired.
- Validation of Suggestions: The study acknowledges that direct validation of its optimization suggestions in the case study was not possible due to a lack of cooperation with local network engineers.
- Future Work:
- Explore semi-supervised learning for perception degradation reasoning, using labels provided by network maintenance engineers.
- Investigate methods to improve the service perception based on the case study's findings, potentially through deeper cooperation with operators.
- Combine crowdsourced data with other sources like Deep Packet Inspection (DPI) and Measurement Reports (MR).
- Leverage high-performance machine learning algorithms for more effective perception degradation positioning.
Summary updated