Final Project : Bigger and Faster Data-graph Computations for Physical Simulations

We investigate the problem of implementing the physical simulations specified in the domain-specific language Simit as a data-graph computation. Data-graph computations consist of a graph G = (V,E), where each vertex has data associated with it, and an update function which is applied to each vertex, taking as inputs the neighboring vertices. PRISM is a framework for executing data-graph computations in shared memory using a scheduling technique called chromatic scheduling, where a coloring of the input graph is used to parcel out batches of independent work, sets of vertices with a common color, while preserving determinism. An alternative scheduling approach is priority-dag scheduling where a priority function ρ mapping each vertex v ∈ V to a real number is used to orient the edges from low to high priority and and thus generate a dag. We propose to extend PRISM in two primary ways. First, we will extend it to use distributed memory to enable problem sizes many orders of magnitude larger than the current implementation using a graph partitioning approach which minimizes the number of edges that cross distributed memory nodes. Second, we will replace the chromatic scheduler in PRISM with a priority-dag scheduler and a priority function which generates a cache-efficient traversal of the vertices when the input graph is locally connected and embeddable in a low-dimensional space. This subset of graphs is important for the physical simulations generated by the language Simit.