We have presented a new IsoDen algorithm for finding halos in N-body cosmology simulations, and described an implementation on parallel computers. This new method has advantages compared to the friends of friends (FOF) algorithm, which has also been implemented in parallel. In particular the IsoDen method robustly finds density peaks even in the high density central regions of clusters. Our tests indicate that these halos are ``real'' in the sense of being gravitationally bound, persistent objects in the simulation, so the IsoDen method is a genuine knowledge discovery process. The use of a statistical estimate of the uncertainty in density estimation to distinguish real peaks from chance associations is novel and effective, even though it lacks a firm theoretical foundation.
By implementing these methods on parallel machines we are able to use them to begin the analysis of the massive datasets produced by modern high resolution N-body cosmology simulations. This will allow us to address the task of accurately interpreting these simulations, to understand the physical processes involved in the formation and evolution of dark matter halos, and to compare the simulations to astronomical observations.
We also consider issues involved in implementing these techniques on inexpensive clusters of commodity processors, using out-of-core techniques to eliminate the need for impractically large amounts of memory. Our implementation of the halo finding techniques in terms of a parallel tree library originally designed for N-body simulation suggests that out-of-core techniques will be practical and viable. A tree library that uses out-of-core techniques is under development, and will form the basis of a parallel, out-of-core implementation of the FOF and IsoDen halo finders. Tracking halos in time, and correlating the results using an ab initio halo finder like IsoDen only at widely separated times will further improve performance.