Doctoral Thesis Defense: Meghrig Terzian
Speaker: Meghrig Terzian
Supervisor: Dr. H. Harutyunyan
Examining Committee: Drs. D. Goswami, B. Jaumard, M. Maheswaran, R. Soleymani, J. Y. Yu (Chair)
Title: On Near Optimal Time and Dynamic Delay and Delay Variation Multicast Algorithms
Date: Friday, December 1, 2017
Time: 10:30 a.m.
Place: EV 11.119
Multicast is one of the most prevalent communication modes in computer networks. A plethora of systems and applications today rely on multicast communication to disseminate traffic including but not limited to teleconferencing, videoconferencing, stock exchanges, supercomputers, software update distribution, distributed database systems, and gaming.
This dissertation elaborates and addresses key research challenges and problems related to the design and implementation of multicast algorithms. In particular, it investigates the problems of (1) Designing near optimal multicast time algorithms for mesh and torus connected systems and (2) Designing efficient algorithms for Delay and Delay Variation Bounded Multicast (DVBM).
To achieve the first goal, improvements on four tree based multicast algorithms are made: Modified PAIR (MPAIR), Modified DIAG (MDIAG), Modified MIN (MMIN), and Modified DIST (MDIST). The proof that MDIAG generates optimal or optimal plus one multicast time in 2-Dimensional (2D) mesh networks is provided. The hybrid version of MDIAG (HMDIAG) is designed, that gives a 3-additive approximation algorithm on multicast time in 2D torus networks. To make HMDIAG applicable on systems using higher dimensional meshes and tori, it is extended and the proof that it gives a (2n-1)-additive approximation algorithm on multicast time in nD torus networks is given.
To address the second goal, Directional Core Selection (DCS) algorithm for core selection and DVBM Tree generation is designed. To further reduce the delay variation of trees generated by DCS, a k-shortest-path based algorithm, Build Lower Variation Tree (BLVT) is designed. To tackle dynamic join/leave requests to the ongoing multicast session, the dynamic version of both algorithms is given that responds to requests by reorganizing the tree and avoiding session disruption. To solve cases where single-core based algorithms fail to construct a DVBM tree, a dynamic three-phase algorithm, Multi-core DVBM Trees (MCDVBMT) is designed, that semi-matches group members to core nodes.