Modeling A Fibre Channel Switch with Stochastic Petri Nets

As computing systems become increasingly complex and concurrent, performance analysis becomes more important, especially when designing an overall architecture. Simple numerical approximations can often yield bounds on performance. Queueing theory and Markovian analysis can provide insight into the steady-state performance of the system. Discreteevent simulation can provide a more detailed view of the system behavior. We used all of the above techniques in the performance study of a Fibre Channel switch. An approximate stochastic Petri net (SPN) model was employed, obtained by exploiting symmetries in the system. The most challenging aspect was determining transition firing probabilities that accurately reflected the dynamic behavior of the exact SPN. This model was valided against a detailed simulation and found to be accurate within 3%. See [1] for details and references. 1 System description We consider a switch with N ports, numbered from 0 to N − 1 (Fig. 1). The switch contains buffer pools organized by input port, a router to transfer header information to output queues, and a scheduler to schedule the transfer of data from an input buffer across the crossbar to an output port. When a message arrives at a port, it is read into a buffer as soon as one is available. The router cyclically polls each port, bypassing ports without new messages. During the poll of a particular port, the router reads the header information of the new message and inserts it into a scheduler table. This table is organized as N FIFO queues, one per output port. The router starts routing a message as soon as the header is i