| Modern NVIDIA’s general-purpose parallel multithreaded graphics processing units (GPU) supplied with the Compute Unified Device Architecture (CUDA) support asynchronous methods through the use of streams. Asynchronous methods can be used to offload great amounts of processing on the GPU to reduce their idle times, resulting in performance increase. In the CUDA context, streams allow simultaneous CPU/GPU operations, concurrent kernel executions, and overlapping of memory transfers with kernel computations. Using a Tesla C2050 GPU the vector addition algorithm is evaluated in parallel synchronous and asynchronous paradigms using CUDA streams. Factors, such as kernel occupancy, kernel decomposition, and memory management affect concurrency in the GPU. The goal of this research is to explore the use of streams as a first stage towards a hybrid framework for concurrent asynchronous algorithms (CAA). CAA relaxes the requirement for synchronization points, the goal being to reduce idle times of all processors. |
