GPUexplore 2.0: Unleashing GPU Explicit-State Model Checking

On the Scalability of the GPUexplore Explicit-State Model Checker

“…3 GPUs and GPUEXPLORE GPUEXPLORE [28,29,33] is an explicit-state model checker that practically runs entirely on a GPU (only the general progress is checked on the host side, i.e. by a thread running on the CPU).…”

“…The general approach of GPUEXPLORE to perform state space exploration is discussed in this sec- Figure 2: Schematic overview of the GPU hardware architecture and GPUEXPLORE tion, leaving out many of the details that are not relevant for understanding the current work. The interested reader is referred to [28,29,33]. In a CUDA program, the host launches CUDA functions called kernels, that are to be executed many times in parallel by a specified number of GPU threads.…”

“…Therefore a single bucket can store 2 * 16 vector length vectors. The reason for writing vectors to buckets with half-warps (a group of 16 threads) is that in many cases, atomic writes of half-warps are scheduled in an uninterrupted sequence [33]. This results in vectors consisting of multiple integers to be written without other write operations corrupting them.…”

“…Accelerating these computations with graphics processing units (GPUs) is one promising way to model check a system design in mere seconds or minutes as opposed to many hours. GPUEXPLORE [28,29,33] is a model checker that performs all its computations on a GPU. Initially, it consisted of a state space exploration engine [28], which was extended to perform on-the-fly deadlock and safety checking [29].…”

“…For instance, the original version running on an NVIDIA K20 was able to explore the state space of the peterson7 model in approximately 72 minutes. With many improvements to GPUEXPLORE's algorithms, reported in [33], the GPU hardware and the CUDA compiler, this has been reduced to 16 seconds. With these levels of speed-up, it has become much more feasible to interactively check and debug large models.…”