Congestion Control in Asynchronous, High-Speed Wormhole Routing Networks

0163-6804/96/$05.00 © 1996 IEEE merging bandwidth-hungry applications (distributed supercomputing, video distribution, visualization of scientific data, etc.) generate increasing demand for costeffective very-high-speed networks. Networks based on linear topologies (e.g., fiber distributed data interface, or FDDI, and distributed queue dual bus, or DQDB) are not suited for veryhigh-speed applications because each transmitter and receiver must operate at the aggregate network speed. Asynchronous transfer mode (ATM) is being widely accepted as the most promising approach to high-speed networking, particularly in the wide area [1]. Local-area ATM is also receiving increasing attention, although costs remain high. Besides cost, a potential limitation of ATM in supercomputer interconnect applications and cluster computing is the high latency, due to bandwidth allocation/negotiation at call setup and cell segmentation and reassembly during the data transfer phase. A promising, cost-effective alternative to ATM for highbandwidth, low-latency applications in the local area and campus environments is represented by the asynchronous wormhole-routing local area network (“worm LAN”). The asynchronous mode of operation allows reduced costs. Variable-size data units (“worms”) make it possible to accommodate different traffic types without the penalty of segmentation overhead. Low latency is achieved by using cut-through or wormhole routing instead of the conventional store-and-forward approach. Wormhole routing has traditionally been the scheme of choice in high-speed networks [2–7] where low-latency data transfer is a key system objective. Wormhole routing with backpressure flow control was originally proposed for multiprocessor computer communications, and later adapted to local-area networking. In wormhole routing, a worm can range in length from a few bytes to several thousands of bytes and has well-defined head, data, and tail portions. When the head of a worm arrives at a node1 and the desired output port is available, the node forwards the worm to the desired output port without waiting for the entire worm to be assembled. Thus, the latency associated with buffering the worm before transmission to the next node is eliminated (whenever possible), reducing the end-to-end latency. With wormhole routing, the worm can stretch across several nodes and links at any one time. The key feature of wormhole-routing high-speed networks is that they can be operated in an asynchronous and unslotted fashion. An example of a commercially available wormholerouting LAN is Myrinet [7]. Myrinet LANs use source routing and 8-bit-wide data transmission to achieve data rates up to 640 Mb/s. Their geographical reach is rather limited, with a maximum link length of 25 m. The Advanced Research Projects Agency (ARPA)-sponsored Supercomputer Super Net (SSN) project at UCLA [8, 9] extends the reach of Myrinet LANs to campus-wide networks via an optical backbone network. Another approach is to replace the bit-parallel copper cables with serial fiber cables. Fiber cable adapters are currently under development at Myricom. Typically, wormhole-routing networks are based on very simple network protocols and therefore are prone to congestion. Congestion is a crucial problem at very high speeds, where large amounts of information can be lost when network resources become unavailable. Some form of flow control must thus be introduced to prevent congestion. A flow control scheme common in worm LANs is backpressure. Backpressure is an explicit link-by-link flow control mechanism requiring bidirectional links. When the desired output port is unavailable, the worm is blocked: while a porEmilio Leonardi and Fabio Neri, Politecnico di Torino