When we analyze the performance of a packet communication network using a queuing model, it is usual to see packets as tokens and to consider that they remain in service for an amount of time proportional to their lengths (the proportionality constant being the reciprocal of the transmission speed). In this paper we show that this paradigm may misrepresent the dynamics of the system, because of its implications on the way memory is used and, in particular, freed. Simulation and analytical results illustrate our claim. Concerning the latter, we provide modified Pollaczec-Khintchine formulas in the M/GI/1 case where we take into account the usual way memory is handled in a communication node.