HTTPDissect: Detailed Performance Analysis of HTTP Web Browsing Traffic in TDMA Mesh Networks

Abstract

TDMA based wireless mesh networks have gained prominence as some of the recent standards such as WiMAX and 802.11s have proposed the use of TDMA based MAC protocol for mesh networks. Even though HTTP web browsing traffic constitutes a significant percentage of Internet traffic, but as of yet there have been no attempts to study the performance of HTTP based web browsing traffic in TDMA mesh networks. HTTP web browsing traffic has different characteristics compared with other types of traffic. In particular, as HTTP traffic consists of large number of small sized file transfers (median file size is 10 KB), it can impose high scheduling overhead. As we highlight, HTTP traffic requires that RTT (round trip time) be small and also it requires that large sized flows be allocated higher share of bandwidth. Given these characteristics of HTTP traffic, it is not clear what protocol design for TDMA mesh networks performs best. In this work, we have studied the performance comparison of four different TDMA MAC protocols for HTTP web browsing traffic. Two of these protocols follow distributed scheduling, one centralized and the other naive static fixed schedule approach. We compare the TDMA MAC protocols for various aspects such as optimum values of different TDMA MAC parameters, design of TDMA scheduling algorithm, performance under varying load, in presence of wireless loss, etc. Our study highlights a crucial aspect that the two distributed protocols specified by the recent WiMAX and 802.11s standards, are incapable of handling wireless packet loss and perform poorly in presence of large HTTP file downloads. As against this, the centralized protocol performs well for the different experiment cases that we have considered. However, our analysis reveals that the performance of the centralized protocol can be further improved if we eliminate the flow set-up phase and minimize the delay in propagating the TDMA schedule. Likewise, we carry out required modifications to the centralized protocol and observe that the performance of the modified protocol, which we term as the sf-centralized (set-up free centralized) protocol, improves by as much as 19 percent under moderate load and 47 percent under high load.