Understanding network performance bottlenecks
MetadataShow full item record
In this thesis we look into network performance bottlenecks and how end-to-end delivery of data is affected by the performance of the network.The last years, the buzzword: the bufferbloat issue has gotten much attention in the network community. Bufferbloat is used to describe problems with high congestion on the Internet. The common wisdom is that congested links and bottlenecks are usually in the last-mile of the network, out to the consumer. This thesis puts this "common wisdom to the test by discovering losses and delays inside the network. A stream of UDP packets were used to measure the end-to-end performance, while the traceroute like tool MTR is used to identify where in the network loss and delay occurs. The experiments were run in a testbed called PlanetLab. PlanetLab is a network of computers distributed across the word. In this work access was gained to servers in most European countries, throughout USA, Argentina, Ecuador, South Korea, Japan, China, Thailand, New Zealand and more. No servers were available in the Middle East, India or Africa.The work is done by designing several Python scripts witch uses MTR and socket programming for sending UDP packet streams. This is done to measure delay and loss in the Internet. The python scripts were deployed and successfully run between over 250 intercontinental pair's servers for one week. The measurements were carefully logged and data was analysed. For the MTR trace-logs, borders between the first- and last-mile networks were established, making it possible to quantify loss and delay occurring in between the identified first-mile hop and the last-mile hop(the transport network). The results from the analysis gave no indication of disproving the common knowledge. Only a low rate of packet loss within the transport-network was discovered. Increased delays, due to congestion, took for the most part place in the Last-mile of the network.