Running time comparisons for the kNN graph computation on GTX 480 (GPU_A) and Tesla C2050 (GPU_B).

The performance showed in terms of running times (in minutes) and speed-ups (x), on four different configurations, single threaded (1 CPU thread), multi-threaded (16 CPU threads), single GPU (GTX 480 (GPU_A) and Tesla C2050 (GPU_B), see Table 1) and multi-GPUs (4 Tesla C2050 GPUs). The time measurements are performed upon repeated executions of the method on each of these data sets and they include the times for loading and transferring data from the to and from the host and device memory. An increase in the chunk size (i.e., increased amount of computations on GPUs) performed better utilization of the parallel hardware and improved the overall the speed-ups. However, the execution times for the single and multi-core CPU implementations remained unchanged with chunk size variations due to the absence of computational chunking and higher chunk sizes could not be applied on GTX480 (GPU_A) due to its limited device memory (1.5 GB approximately). A total of 295 samples and a single value for () are used to perform these tests.