Time-based performance evaluation of protocols for distributed systems: An alternative to message counting

The development of many protocols for distributed systems is based on a performance metric which is the asymptotic worst-case number of messages required to implement the protocol. It is shown that such an analysis may result in false conclusions about the relative performance of competing protocols. A performance evaluation of the logical timestamp protocol (LTP) and the immediate ordered service protocol (IOS) for distributed first-come-first-served service is presented, using GENSIM, a general purpose facility for simulating distributed systems. Expected-case analysis of time-based metrics may be quite pessimistic when a high degree of concurrent activity is possible in the system. The use of GENSIM to structure a set of simulations procedures metrics which indicate that IOS is superior to LTP for most reasonably small and reasonably fast networks is discussed. The performance metric of choice is real time, a natural metric which is easily handled by simulation but which is only tangentially related to message counts. >