Model Checking via Delayed Duplicate Detection on the GPU

In this paper we improve large-scale disk-based model checking by shifting complex numerical operations to the graphic card, enjoying that during the last decade graphics processing units (GPUs) have become very powerful. For disk-based graph search, the delayed elimination of duplicates is the performance bottleneck as it amounts to sorting large state vector sets. We perform parallel processing on the GPU to improve the sorting speed significantly. Since existing GPU sorting solutions like Bitonic Sort and Quicksort do not obey any speed-up on state vectors, we propose a refined GPU-based Bucket Sort algorithm. Alternatively, we study sorting a compressed state vector and obtain speed-ups for delayed duplicate detection of more than one order of magnitude with a single GPU, located on an ordinary graphic card.