Effective Wide-Area Network Performance Monitoring and Diagnosis from End Systems.

Abstract

The quality of all network application services running on today’s Internet heavily depends on the performance assurance offered by the Internet Service Providers (ISPs). Large network providers inside the core of the Internet are instrumental in determining the network properties of their transit services due to their wide-area coverage, especially in the presence of the increasingly deployed real-time sensitive network applications. The end-to-end performance of distributed applications and network services are susceptible to network disruptions in ISP networks. Given the scale and complexity of the Internet, failures and performance problems can occur in different ISP networks. It is important to efficiently identify and proactively respond to potential problems to prevent large damage.

Existing work to monitor and diagnose network disruptions are ISP-centric, which relying on each ISP to set up monitors and diagnose within its network. This approach is limited as ISPs are unwilling to revealing such data to the public. My dissertation research developed a light-weight active monitoring system to monitor, diagnose and react to network disruptions by purely using end hosts, which can help customers assess the compliance of their service-level agreements (SLAs). This thesis studies research problems from three indispensable aspects: efficient monitoring, accurate diagnosis, and effective mitigation. This is an essential step towards accountability and fairness on the Internet.

To fully understand the limitation of relying on ISP data, this thesis first studies and demonstrates the monitor selection’s great impact on the monitoring quality and the interpretation of the results. Motivated by the limitation of ISP-centric approach, this thesis demonstrates two techniques to diagnose two types of finegrained causes accurately and scalably by exploring information across routing and data planes, as well as sharing information among multiple locations collaboratively. Finally, we demonstrate usefulness of the monitoring and diagnosis results with two mitigation applications. The first application is short-term prevention of avoiding choosing the problematic route by exploring the predictability from history. The second application is to scalably compare multiple ISPs across four important performance metrics, namely reachability, loss rate, latency, and path diversity completely from end systems without any ISP cooperation.