All-optical WDM networks are fast becoming the natural choice for future backbone. In this paper, we establish the efficiency of multicasting over unicasting in all-optical WIDM networks, assess the usefulness of wavelength conversion for multicasting, and explore the issues related to the splitting (or copying) capability of the nodes. The comparison between multicasting and unicasting is based on the number of wavelengths as well as the amount of bandwidth required for a given set of multicasting sessions. For each multicasting session, a source-specific multicasting forest (or trees) is constructed first, taking into account the sparse splitting capability of the nodes in the network. Then, each multicasting tree is partitioned into segmenis according to the sparse wavelength conversion capability of the nodes on the tree such that each segment needs to be assigned the same wavelength. Simulation results obtained for a practical network such as NSFNET and randomly generated networks show that multicasting can reduce both the bandwidth consumed and the number of wavelengths required by as much as 50% or more when the size (i.e. the number of destinations) of each multicasting session is reasonably large. Such a reduction due to multicasting is not affected much by the wavelength conversion capability, the number of multicasting sessions and the size of the networks whose topology is more or less random. The results have also shown that sparse splitting can be nearly as effective as full splitting for multicasting.