A set of performance criteria for evaluating optimization software with respect to efficiency, reliability, and accuracy is presented and discussed. A numerical comparison of five constrained nonlinear programming codes is described, which was carried out in order to test the usefulness and general applicability of the proposed performance criteria. The results of the numerical comparison are discussed, and the proposed criteria are compared to the criteria traditionally used in comparative evaluations of nonlinear programming codes, with particular reference to machine dependence and the applicability to test problems with unknown solutions. A separate small scale computational experiment is described which was carried out specifically to test the machine dependence of the criteria. The observed deficiencies of the proposed new criteria are also discussed.