Assessment of Burst Assembly Algorithms using Real IPv4 Data Traces

Abstract

This paper describes the generation of data bursts using discusses the metrics devised for assessment of the burst real JPv4 data as input, compares the performance of the three assembly algorithms; Section IV presents the assessment of the aggregation algorithms and suggests a novel mechanism to adjust the burst assembly algorithms and Section V presents the novel burst assembly algorithm thresholds. The generated bursts are mechanism to adjust the burst assembly algorithm thresholds studied, in particular average packet delay per burst, burst interand main conclusions.

Issues on Performance Assessment of Optical Burst Switched Networks: Burst Loss Versus Packet Loss Metrics

“…This also points out to conclusions already known from the study of burst assembly algorithms using real IP packets: minimum packet delay and maximum burst size, i.e. optimization of burst assembly process, is achieved for the HA algorithm using time and size thresholds that are function of the network load on the burst assembly machine, and thus, are network point dependent [8]. As a result of the optimization of the burst assembly process, it has to be assumed that realistic burst switching deals with bursts that are not homogeneously sized, and of course do not contain a fixed number of packets [8].…”

“…2 shows how different sets of thresholds change the inter-arrival time between bursts, and consequently define the optimum zone for burst assembly algorithms, defined as corresponding to the minimum interarrival time between bursts with the maximum burst size. Since burst assembly thresholds are network point dependent [8], HA was used with a wide set of thresholds as to obtain a large range of burst characteristics. The result was the creation of bursts very differentiated in terms of Size (in Bytes) and Size (in number of Packets), results that are clearly visible in Fig.…”

“…optimization of burst assembly process, is achieved for the HA algorithm using time and size thresholds that are function of the network load on the burst assembly machine, and thus, are network point dependent [8]. As a result of the optimization of the burst assembly process, it has to be assumed that realistic burst switching deals with bursts that are not homogeneously sized, and of course do not contain a fixed number of packets [8]. If the three above mentioned assumptions can not be held true, as in the case where very heterogeneous burst traffic is generated (the case simulated and presented here), only two alternatives remain: either burst loss is not adequate as a performance assessment metric because it is not equal neither to byte loss neither to packet loss, and the latter would be more "user meaningful", or with real traffic the simulation proves the Law of Big Numbers, and so, the final results on the network can be assumed as if all the bursts have the same number of packets, and these in turn are equally sized, to the average number of packets per burst the first, and the average number of bytes per packet (and per burst) the latter.…”