Retransmission policies for multihomed transport protocols

Abstract

We evaluate three retransmission policies for transport protocols that support multihoming (e.g. SCTP). The policies dictate whether retransmissions are sent to the same peer IP address as the original transmission, or sent to an alternate peer IP address. Each policy presents tradeoffs based on the paths' bandwidth, delay, loss rate, and IP destination reachability. We find that sending all retransmissions to an alternate peer IP address is useful when the primary IP address becomes unreachable, but often degrades performance in non-failure scenarios. On the other hand, sending all retransmissions to the same peer IP address as the original transmission reverses the tradeoffs. We balance the tradeoffs by proposing a hybrid policy that sends fast retransmissions to the same peer IP address as the original transmission, and sends timeout retransmissions to an alternate peer IP address. We show that even with extensions which we proposed to improve the policies' performance, the hybrid policy is the best performing policy in failure and non-failure scenarios. q

Multistreamed web transport for developing regions

“…Unlike TCP whose SACK info is limited by the space available for TCP options, the size of SCTP's SACK chunk is limited by the path MTU and contains more accurate information about lost TPDUs than TCP. Also, FreeBSD's SCTP stack implements the Multiple Fast Retransmit algorithm (MFR), which reduces the number of timeout recoveries at the sender [Caro 2006]. Therefore, as loss rates increase, we expected the enhanced loss recovery features to help SCTP perform better than TCP.…”

A taxonomy and survey of SCTP research

“…The retransmission policy, that is, on which path the retransmissions should be sent and under which constraints, has been extensively studied by Caro et al [2003aCaro et al [ , 2003bCaro et al [ , 2004bCaro et al [ , 2006b]. Three basic retransmission policies were evaluated: send all retransmissions on the alternate path, send all retransmissions on the primary path, and a hybrid policy that sends fast retransmissions on the primary path and timeout retransmissions on the alternate path.…”

“…Three basic retransmission policies were evaluated: send all retransmissions on the alternate path, send all retransmissions on the primary path, and a hybrid policy that sends fast retransmissions on the primary path and timeout retransmissions on the alternate path. The earlier studies [Caro et al 2003a[Caro et al , 2003b[Caro et al , 2004b] use a symmetric configuration in which both paths have the same bandwidth and delay characteristics, whereas the later study [Caro et al 2006b] considers also the case of path asymmetry. After their comprehensive evaluation of both failure and non-failure scenarios, the recommendation of Caro et al is to adopt the hybrid retransmission policy that sends fast retransmissions to the same IP address as the original transmission and sends timeout retransmissions to an alternate peer IP address.…”

Rethinking end-to-end failover with transport layer multihoming

“…On the other hand, uniform loss is a simple, yet sufficient model to provide insight about the effectiveness of different PMR settings accurately detecting failure. To evaluate if Figure 2's loss model was reasonable, we compared representative simulations using a cross-traffic model, shown in [11], to produce self-similar, bursty traffic. Although the absolute results differed for those examples compared, relative relationships remained consistent -leading to the same conclusions.…”