The authors consider the queue arising in a multiservice network using ATM (asynchronous transfer mode) when a superposition of periodic streams of constant-length cells is multiplexed on a high-speed link. An exact closed formula is derived for the queue length distribution in the case where all streams have the same period, and tight upper and lower bounds are obtained on this distribution when the periods are different. Numerical results confirm that the use of a Poisson approximation (i.e. the M/D/1 queue) can lead to a significant overestimation of buffer requirements, particularly in the case of heavy loads. Buffer requirements for a mixture of different period streams can be accurately estimated from the upper bound on the queue length distribution. For given load, requirements increase with the number of long-period (i.e. low-bit rate) sources. The results are deduced from a novel characterization of the single-server constant service time queue, which should be useful in other applications.