Recently RFID technology has made its way into end-user applications, enabling automatic item identification without requiring line of sight. In particular passive tags provide a promising, low cost and energy-efficient solution for inventory applications. However, their large-scale adoption strictly depends on the efficiency of the identification process. A major challenge is how to arbitrate channel access so that all tags are able to answer the reader inquiries and identify themselves over time. This paper stems from the observation that a variety of anti- collision protocols for RFIDs have been proposed in the literature. However, a thorough simulation comparison among them and a clear identification of the mechanisms resulting in better end- to-end performance is lacking. The objective of our work has been to fill this gap. This paper presents the results of a detailed ns2-based comparative evaluation of representatives of all the classes of anti-collision protocols so far proposed. Simulation results show that end-to-end performance of the different classes of protocols in terms of metrics such as the time needed for tags identification differ significantly over what previously found by experiments which only focused on the number of reading cycles for tag identification. Our thorough performance evaluation has highlighted that different solutions are to be used in different application scenarios and that decreasing the collisions (rather than idle times) is the way to go to further improve anti-collision protocols performance. I. INTRODUCTION A basic RFID system consists of a reader and a set of tags. The reader inquiries tags that are able to communicate on the radio channel, returning their ID. Tags are typically passive devices, which answer to reader's query by back- scattering the received signal. One of the main objectives of a RFID system is the identification of all the tags present in the area covered by the reader. The challenges related to tags identification depend on the reference scenarios, which may include one or more readers, and a variable or stationary set of tags. The coexistence of multiple readers in the same area may cause collisions among readers interfering with each other or with tags. Moreover, collisions may occur among tags simultaneously transmitting to the same reader, indepen- dently of the presence of one or more readers. Collisions are addressed by specific solutions for the multi-reader problem (through frequency allocation mechanisms) and the single- reader problem (through collision arbitration schemes). The other distinguishing factor is given by scenario variability. In case of stationary applications scenarios (i.e., consecutive readings of the same, slightly changed, set of tags), the reader