Terabit Burst Switching Progress Report (9/98–12/98)

This report summarizes progress on Washington University’s Terabit Burst Switching Project, supported by DARPA and Rome Air Force Laboratory. This project seeks to demonstrate the feasibility of Burst Switching, a new data communication service which can more effectively exploit the large bandwidths becoming available in WDM transmission systems, than conventional communication technologies like ATM and IP-based packet switching. Burst switching systems dynamically assign data bursts to channels in optical data links, using routing information carried in parallel control channels. The project will lead to the construction of a demonstration switch with throughput exceeding 200 Gb/s and scalable to over 10 Tb/s. This work is supported by the Advanced Research Projects Agency and Rome Laboratory (contract F30602-97-10273). Terabit Burst Switching Progress Report (9/98–12/98) Jonathan S. Turner [email protected] This report summarizes progress on the Terabit Burst Switching Project at Washington University for the period from September 15, 1998 through December 15, 1998. 1. Improved Link Scheduling One of the key issues in the design of a burst switch is scheduling the transmission of bursts on outgoing channels. Previously, we have concentrated on a technique called horizon scheduling. In horizon scheduling, the link scheduler maintains a time horizon for each of the channels of an outgoing link. The horizon is defined as the earliest time after which there is no planned use of the channel. The horizon scheduler assigns arriving bursts to the channel with the latest horizon that is earlier than the arrival time of the burst, if there is such a channel. If there is no such channel, the behavior depends on whether the system supports burst storage or not. In a system that does not support burst storage, the scheduler simply discards the burst. In a system that does support storage, the burst is assigned to the channel with the smallest horizon and is delayed until that channel becomes available. Horizon scheduling is straightforward to implement in hardware, but because it does not keep track of time periods before a channel’s horizon when the channel is unused, it cannot insert bursts into these open spaces. Figure 1 illustrates the operation of a horizon scheduler. The diagram shows an arriving burst that needs to be assigned to an outgoing channel. The burst header cell arrives at time t with the start of the burst arriving at t+∆1 and the end of the burst arriving at t+∆2. The right hand part of the figure shows bursts that arrived earlier that were assigned to channels by the horizon scheduler. The thick lines indicate time periods during which the various channels are scheduled to be in use and the shaded region shows those time periods that are currently unavailable to new bursts. The box in the center highlights the time period during which the arriving burst will need an outgoing channel and the check marks at the right indicate those channels that it could be assigned to. Since the horizon scheduler prefers the viable channel with the latest horizon, it will select the bottom channel. In some situations, the horizon scheduler can provide good performance. Let b1; : : : ; bn be a sequence of bursts, where bi is characterized by a triple (ri; ti; `i) and ri is the time at which the link scheduler receives the burst header cell informing it of the imminent arrival of the burst, ti is the