Modeling GPU Dynamic Parallelism for Self Similar Density Workloads

Dynamic Parallelism (DP) is a runtime feature of the GPU programming model that allows GPU threads to execute additional GPU kernels, recursively. Apart from making the programming of parallel hierarchical patterns easier, DP can also speedup problems that exhibit a heterogeneous data layout by focusing, through a subdivision process, the finite GPU resources on the sub-regions that exhibit more parallelism. However, doing an optimal subdivision process is not trivial, as there are different parameters that play an important role in the final performance of DP. Moreover, the current programming abstraction for DP also introduces an overhead that can penalize the final performance. In this work we present a subdivision cost model for problems that exhibit self similar density (SSD) workloads (such as fractals), in order understand what parameters provide the fastest subdivision approach. Also, we introduce a new subdivision implementation, named \textit{Adaptive Serial Kernels} (ASK), as a smaller overhead alternative to CUDA's Dynamic Parallelism. Using the cost model on the Mandelbrot Set as a case study shows that the optimal scheme is to start with an initial subdivision between $g=[2,16]$, then keep subdividing in regions of $r=2,4$, and stop when regions reach a size of $B \sim 32$. The experimental results agree with the theoretical parameters, confirming the usability of the cost model. In terms of performance, the proposed ASK approach runs up to $\sim 60\%$ faster than Dynamic Parallelism in the Mandelbrot set, and up to $12\times$ faster than a basic exhaustive implementation, whereas DP is up to $7.5\times$.