Proposition and simulation study of a novel heavy hitter detection mechanism

Abstract

In the present dissertation is proposed a new mechanism with three (3) implementable variations (counting schemes) for the detection of the most significant so called Heavy Hitters (flows whose load inside a flow-set is very relevant) and or a group of them. The mechanisms is to be applied in egress router queues; it consist in evidencing every certain period of time (called HCP) what is considered as the HHs applying some counting scheme. Moreover, inside each HCP is done a fix number of consecutive queue checks separated by a time interval; at every check, is count the number of packets for each of the found flows and the information of all those checks is related by the counting scheme; that information is only valid during one HCP. The counting schemes are: (1) cumulative, (2) vector, (3) mean. The mechanism is evaluated in a controlled simulation scenario in where can be found TCP and UDP flows with a variety of bit-rates. In particular is denoted the attention to the cumulative scheme which in implementation considerations shows to the fastest and less memory consuming of the schemes. In the study was determined that increasing the number of queue checks improves the selection of the HHs by the mechanism; on the other side, increasing the time interval between checks has not much effect in the selection. Where compared the results of the Cumulative-CS with those offered by the Mean-CS and was found that the performance in finding the most relevant flows was better for the cumulative counting scheme which is also the fastest and less memory consuming of the schemes.