| Over the years, different architectures have been investigated for the design and implementation of high-performance switches. Particular architectures were determined by a number of factors based on performance, flexibility and available technology. Design differences were mainly a variation in the queuing functions and the switch core. The crossbar-based architecture is perhaps the dominant architecture for today’s high-performance packet switches (IP routers, ATM switches, and Ethernet switches) and owes its popularity to its scalability (when compared to the shared-bus/shared-memory architectures), efficient operation (supports multiple I/O transactions simultaneously) and simple hardware requirements. The architecture includes the input-queued (IQ) crossbar fabric switch with its variations (Outputqueued, OQ, switch and Combined Input–Output-queued, CIOQ, switch) and the internally buffered crossbar fabric switch (BCS).
IQ switches have gained much interest in both academia and industry because of their low cost and scalability. The IQ switch has a low internal speedup because the crossbar fabric has the same speed as that of the external line. Although the headofline (HoL) blocking problem limits the achievable throughput of an IQ switch to approximately 58.6% [1], the well-known virtual output queuing (VOQ) architecture [2] was proposed and has improved switching performance by several orders of magnitude, making IQ switches more desirable. However, the adoption of VOQ has created a more serious problem, namely, the centralized scheduler. An arbitration algorithm examines the contents of all the input queues, and finds a conflictfree match between inputs and outputs. The well-known optimal algorithms (i.e. maximum-weight-matching or MWM) are too complex to implement at high speed while the iterative algorithms, proposed as an alternative to the MWM algorithms, fail to perform well under real world input traffic conditions. |