Modeling GPU Dynamic Parallelism for self similar density workloads

Dynamic Parallelism (DP) is a GPU programming abstraction that can make parallel computation more efficient for problems exhibit heterogeneous workloads. With DP, threads launch kernels with threads, recursively, producing subdivision effect where resources are focused on the regions work. Doing an optimal process not trivial, as combination of different parameters play relevant role in final performance DP. Also, current DP relies kernel recursion, which has overhead. This work presents new cost model self similar density (SSD) workloads, useful finding schemes. implementation free recursion overhead presented, named Adaptive Serial Kernels (ASK). Using Mandelbrot set case study, shows achieved when using {g∼32,r∼2,B∼32} initial subdivision, recurrent and stopping size, respectively. Experimental results agree theoretical parameters, confirming usability model. In terms performance, ASK approach runs up to ∼60% faster than set, 12× basic exhaustive implementation, whereas 7.5× faster. energy efficiency, ∼2× ∼20× approach, These put tools analyzing potential improvement based approaches developing GPU-based libraries or fine-tune specific codes research teams.