The International Conference for High Performance Computing, Networking, Storage and Analysis
Performance Grading of GPU-Based Implementation of Space Computing Systems Image Compression.
Authors: Olympia Kremmyda (National and Kapodistrian University of Athens), Vasilis Dimitsas (National and Kapodistrian University of Athens), Dimitris Gizopoulos (National and Kapodistrian University of Athens)
Abstract: Modern GPUs can accelerate the execution of inherently data-parallel applications. However, real-world algorithm realizations on GPUs differ from synthetic benchmarks. In this paper we stress the performance limits of GPUs. We measure the performance of our CUDA implementation of a recommended image compression algorithm (CCSDS-122.0-B-1) for space data systems on NVIDIA GPUs for various image sizes and compare it to the execution on x86 CPUs. The control-intensive parts and irregular memory access patterns under-utilize the GPU architecture and effectively serialize large parts of the GPU kernel execution.
We present the execution results on two systems with different CPU/GPU setups. In both systems, our GPU implementation of the algorithm eventually leads to a very moderate 1.1x to 2.1x speedup. Discussion of the performance bottlenecks leads to insights on how GPU execution can be improved in the future by either revising the algorithms or by supporting control-intensive execution on the GPU cores.