A fine-grained response time analysis technique in heterogeneous environments

It is crucial for the network operators and Internet service providers (ISPs) to determine the reasons that cause large response time fluctuations. In this paper, we consider passive measurements from heterogeneous environment (ADSL, FTTH and 3G/3G+ access technologies) of an European ISP 'Orange'. Through experimental analysis of real traces, the need of a fine-grained traffic analysis technique is demonstrated. We show that finding the root causes of the observed poor performance using simple metrics such as response time, RTT and packet loss is difficult. In view of this fact, the different factors that play a role in determining the resulting response time are described through examples. Then, a breakdown method that drills down into the passively observed TCP connections is proposed. The method decomposes the end-to-end response time into many time periods and maps each one to a specific parameter or a physical phenomenon. Thus, the impact of not only the network parameters but also the application configuration and user behavior is captured. The resulting time periods are given as input to a clustering algorithm in order to group together transfers with similar performance holding traffic of different application protocols over different access technologies. As a result, the contribution of each participant in the performance bottleneck is identified. The proposed technique is validated through extensive simulations and real passively measured traces and it is compared to other works. Exemplifying the technique on real traces from Internet and enterprise traffic is introduced and discussed to demonstrate the power of the approach and its simplicity. In contrast to some existing tools, ISPs and enterprise administrators do not need to modify their network architecture or to install a new software or a plugin at the client or at the server side in order to use our technique. In addition, data sampling is not used. This is particularly important in order to keep data consistency and to detect metrics peaks. Last, our tool deals with both long and short TCP connections.