Performance implications of a bounded receive buffer in concurrent multipath transfer
Computer Communications
Employing transport layer multi-railing in cluster networks
Journal of Parallel and Distributed Computing
Evaluating TCP-friendliness in light of Concurrent Multipath Transfer
Computer Networks: The International Journal of Computer and Telecommunications Networking
Using CMT in SCTP-based MPI to exploit multiple interfaces in cluster nodes
PVM/MPI'07 Proceedings of the 14th European conference on Recent Advances in Parallel Virtual Machine and Message Passing Interface
Hi-index | 0.00 |
Transport layer multihoming binds a single transport layer association to multiple network addresses at each endpoint, thus allowing the two end hosts to communicate over multiple network paths. This dissertation investigates end-to-end Concurrent Multipath Transfer (CMT) using transport layer multihoming for increased application throughput. CMT is the simultaneous transfer of new data from a source host to a destination host via two or more end-to-end paths. We investigate and evaluate design considerations in implementing CMT at the transport layer using the Stream Control Transmission Protocol (SCTP) as an example of a multihome-capable transport layer protocol. Specifically, we explore (i) algorithms for CMT at the transport layer, (ii) retransmission policies for CMT, and (iii) performance implications of a bounded receive buffer on CMT. We identify three negative side-effects of reordering due to CMT that must be managed before the full performance gains of CMT's parallel transfer can be achieved. We propose three algorithms to eliminate these side-effects: the Split Fast Retransmit algorithm (SFR) to handle unnecessary fast retransmissions by a sender, the Cwnd Update for CMT algorithm (CUC) to counter overly conservative congestion window growth at a sender, and the Delayed Ack for CMT algorithm (DAC) to curb an increase in ack traffic due to fewer delayed acks by a receiver. These algorithms demonstrate that a single sequence space within a transport layer association is sufficient for CMT; separate sequence spaces per path are not required.We propose and evaluate five retransmission policies for CMT. Introducing these retransmission policies causes two side-effects: occurrence of spurious retransmissions and inaccurate congestion window estimation. We propose two protocol modifications to eliminate these side-effects. Using simulation, we evaluate CMT against AppStripe, a simulated idealized application that stripes data over multiple paths using multiple SCTP associations. The results of this evaluation demonstrate that CMT's sharing of sequence space across paths improves performance---an inherent benefit that load sharing at the transport layer has over that at the application layer. We study the performance of CMT in the presence of a bounded receive buffer (rbuf). Simulation results show that if two paths are used for CMT, the lower quality (i.e., higher loss rate) path degrades overall throughput of an rbuf-constrained CMT association by blocking the rbuf. (Abstract shortened by UMI.)