Dept. of Computer Science and Engineering
Oregon Graduate Center
The Independent Time Step Method (ITSM) exploits the underlying parallelism in physics simulations by using local information to advance physical values in time. Spatial regions containing physical data are allowed to update asynchronously. Regions communicate with near neighbors, and mutual exclusion is required to assure consistency of updates. Also, the neighborhood of a region can change with time. ITSM is an example of a tightly coupled and relatively fine grained parallel algorithm based on an arbitrary dynamic graph. Early attempts at parallelization of ITSM codes focused on allocation of work to processes. This was the starting point for this study of alternative scheduling and synchronization strategies for these codes, using models that incorporated the major features of the ITSM problem space. Experiments measuring the performance of the models over a range of parameter settings were carried out on a Sequent Symmetry. The models are differentiated based on the strategy a region uses when blocked by an executing neighbor, and by the time at which synchronization occurs. The performance of each model is discussed, and the results are compared. Results are also given for a full ITSM code based on a version of one of the models. This code achieved linear speedup.
Storc, Lise, "Parallelization schemes for physics codes using the independent time step method" (1989). Scholar Archive. 148.