Performance evaluations and comparisons of mutual exclusion algorithms have been mainly focused on algorithm analyses in terms of their computational complexities. However, effects from architectures and implementation variations also play important roles in performance of the algorithms. In this article, we propose a framework for mutual exclusion algorithm performance evaluation on shared-memory multiprocessors, which combines three important factors, namely, the properties of the algorithms, execution complexities, and architecture and system effects. We implemented several representative mutual exclusion algorithms on two different shared-memory multiprocessor systems: the BBN TC2000 and the KSR-1. Using measured network latencies of algorithm executions, we evaluate and compare the scalabilities of the algorithms. Finally, we propose new and efficient mutual exclusion algorithms that combine good features of existing mutual exclusion algorithms. Our results show that execution performance of software-based algorithms are highly architecture-dependent. In general, a hardware-based algorithm performs more efficiently than a software-based algorithm in the presence of contention, due to support from atomic instructions. However, software-based algorithms have a unique advantage of flexibility, which can be used to take advantage of the structure of architectures and systems. Execution performance of these new software-based algorithms shows their potential to rival against hardware-based algorithms for mutual exclusion on shared-memory architectures.