The ready-to-go virtual circuit protocol: a loss-free protocol for multigigabit networks using FIFO buffers

The ready-to-go virtual circuit protocol (or RGVC) is an immediate transmission protocol, in which the source need not wait for an end-to-end roundtrip delay for reservations to be made before transmitting the data. The protocol is designed to handle the lossless transport of ABR traffic, and will be used in the 40 Gb/s Thunder and Lightning testbed being prototyped at the University of California at Santa Barbara (UCSB). An important advantage of the RGVC protocol over previous connection and flow control protocols is that it is suitable for networks in which the switches use FIFO buffers that are shared by multiple sessions. The RGVC protocol ensures lossless communication by coupling link capacity with buffer space, so that when a portion of a buffer at a node is occupied, a proportional fraction of the incoming capacity to that buffer is frozen. Given the constraints on the frozen capacity, an algorithm is executed at each node to allocate the transmission rate to each FIFO buffer so as to maximize capacity utilization. The requirement that the protocol operate with FIFO buffers at the network nodes poses some unique challenges in the design that are not present in rate- and credit-based schemes. Briefly, since several sessions share a common FIFO buffer, per-VC flow control is no longer possible so control over the rate of an individual session is lost. Also, since the contents of the buffers change dynamically, the buffer composition becomes difficult to determine. For the rate-allocation algorithm of the RGVC protocol to be executed, however, the contents of the FIFO buffers at a node must be known, To implement the bookkeeping required, we present two schemes: the measurement-based scheme, where the bookkeeping function is implemented via measurements, done essentially in hardware; and the estimation-based scheme, where the bookkeeping is done analytically via the exchange of control packets between nodes.