A methodology is presented for the performance evaluation of large stochastic high—level Petri net(SHLPN)models that are structured into independent submodels．These subnets are independently evaluated and then are substituted by the nonprimitive transitions as performance equivalence in the original model to achieve a state space reduc tion．This is based on the hierarchical design of SHLPN models and on the estimed mean delay time for a token traffic process in subnets．The paper investigates four subnet inter-faces for decomposition and composition as performance and computation of the mean response time for subnets．