Design and implementation of parallel list data structure using graphics accelerators

One of the biggest shortcomings of the CUDA environment is the lack of a data structure having effective parallel insert and remove operations. This paper focuses on the insertion. The traditional vector-based lists are not applicable for this kind of addition; however, the well-known linked list data structure has the ability to handle parallel insertions at the same time. But nowadays, there is not any “official” linked list data structure in the current CUDA runtime library. This paper presents a novel way to create a multi-layered linked list (called Parallel List) using the NVIDIA CUDA framework. As our experiments show, the implemented data structure is able to do multiple insert operations 2-30 times faster than the traditional sequential CPU implementation of a general list object. Another advantage of the new data structure is the support of Random Access pattern to access contents directly.