טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentNebat Yoav
SubjectResequencing Considerations in Downloads
DepartmentDepartment of Electrical Engineering
Supervisors Professor Emeritus Raphael Rom
Professor Moshe Sidi


Abstract

Several recent studies have proposed methods to accelerate the receipt of a file by downloading its parts from different servers in parallel. Note that with this approach data segments are likely to arrive out-of-order, which may cause large resequencing-buffer demands, and re-ordering, resulting in an unregulated flow of data to the application at the destination. As recent papers indicate, when unicast TCP is used, the occurring of reordering may result in unfavorable consequences, including self-reinforcingly poor performance.

We present models formulated for an approach based on receiving only one copy of each of the data packets in a file, while different packets may be obtained from different sources. This approach guarantees faster downloads with lower network use.

Methods to derive the probability distribution of the resequencing-buffer occupancy for several possible scenarios are presented. This is important for the determination of necessary buffer sizes at the receiver to keep potential overflows

under a specific threshold. The resequencing-buffer occupancy is also a good indicator to the severeness of out-of-order arrival.

Algorithms devised to keep out-of-order low to insure a regulated flow of packets to the application will be presented. The basic idea is to obtain, before or during the reception process, statistics on the current rate of transmission for all the relevant sources and on the delays on the particular path that each source will use for transmission to the user. These parameters can be obtained by applying an end-to-end measurement method. Based on that information, specific packets are demanded from each source for fast reception of the entire file and for low reception time variance. Performance analysis of these algorithms is also presented.