Browsing by Author "Wenke Lee, member"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • No Thumbnail Available
    Adaptive Multipath Traffic Allocation in TCP/IP Networks
    (2001-09-05) Elhaddad, Mahmoud Shawky; Munindar Singh, Chair; George Rouskas, member; Wenke Lee, member
    Balancing network traffic among multiple paths connecting ingress-egress pairspromises better utilization of network links and improved servicequality for user-flows.Research into optimized adaptive traffic allocation yielded techniquesthat tradeoff convergence speed to stability and/or fail to incorporate knowledgeabout the behavior of congestion control mechanisms, making their evaluationa difficult, if possible task -- both analytically and experimentally.In this thesis, we consider load balancing in multipath IP networks where thenumber of flows between ingress-egress pairs is a slow-changing process andTCP-friendliness is adopted-by or imposed-on all flows. Provided that packetordering is preserved, we present an analytical characterization of the throughput-optimalfractional allocations of flow packets in terms of the TCP-fair share alongeach candidate path; and show that optimally splitting all network flows resultsin efficient and globally fair sharing of network resources. For the estimationof the fair share along network paths, we demonstrate that TCP-fairness in sharingbottleneck bandwidth can be modeled as weighted max-min, where the weight vectorcorresponds to the Bulk Transfer Capacity (BTC) of the bottlenecked path. Policingat the ingress routers limits connections to their computed fair rates, therebyeliminating TCP bias against connections passing through multiple bottlenecksand minimizing packet losses at internal routers. Given minimal loss rate andsmall bounded delay skew among alternative paths, packet resequencing at theegress routers can be applied effectively.We also introduce AMTA -- a centralized traffic allocation service for MPLSnetworks based on the optimality and fairness results described above, and extendedto handle flows with limited throughput demand. The robustness of AMTA underrelevant scenarios is demonstrated through simulations.