Benchmarks Based on Anti-Parallel Patterns for the Evaluation of GPUs

We put forward “anti-parallel patterns” to guide the parallel performance analysis process. Anti-parallel patterns or APPs are common parts of parallel programs that cause these programs to have less than ideal performance, where the ideal speedup equals the number of processors. We present benchmarks to model the behavior of APPs on parallel platforms. Each benchmark contains only one APP and is configurable to mimic all its instances. We show how benchmarks can be used to qualitatively and quantitatively understand and compare parallel hardware. Experiments with NVIDIA and AMD GPUs reveal their differences.