Cell loss, which occurs when buffers overflow in ATM switches, is a ke
y measure for the performance of an ATM network. An ATM network must b
e designed for very low and predictable cell loss probabilities to pro
vide adequate quality of service for all users. Many methods have been
proposed for allocating bandwidth and determining buffer size require
ments to provide quality-of-service guarantees. The probabilistic natu
re of these methods is subject to overly optimistic allocations (insuf
ficient in the worst case) or overly pessimistic allocations (over spe
cified in the best case). The authors present an exact analysis of cel
l loss for periodic on/off (bursty) traffic sources. Periodic on/off t
raffic sources are representative of output traffic from leaky-bucket
rate-control mechanisms as used in ATM networks. Closed-form equations
for buffer sizing and cell loss probabilities are developed and rigor
ously proved for a single queue with multiple traffic sources and mult
iple queues with multiple traffic sources serviced via a round-robin d
iscipline. These closed-form equations can be used in concert with pro
babilistic methods to ensure that buffer and bandwidth allocations are
made between the best- and worst-case requirements.