Časově deterministické vykonávání kódu na GPU

Abstract

In this thesis, we evaluate the interference between multiple GPU (Graphics processing unit) kernels running in parallel based on artificial random and sequential walks, the 2D Convolution benchmark provided by polybench and the KCF (Kernelized Correlation Filter) tracker implemented by Vít Karafiát and Michal Sojka. To achieve a reduction of the interference between the running kernels and to reduce the resulting execution jitter, we used a time-triggered execution on the GPU. To enable the synchronization, we assessed two synchronization mechanisms available on Tegra X2 platform: one based on zero-copy memory and one based on the globaltimer. We found that the NVIDIA profiler (nvprof) reconfigures the resolution of globaltimer from 1 s to 160 ns. With this resolution, we were able to reduce the execution time jitter of a tiled 2D convolution kernel from 6.47% to 0.15% while maintaining the same average execution time by use of a time-triggered GPU execution.

In this thesis, we evaluate the interference between multiple GPU (Graphics processing unit) kernels running in parallel based on artificial random and sequential walks, the 2D Convolution benchmark provided by polybench and the KCF (Kernelized Correlation Filter) tracker implemented by Vít Karafiát and Michal Sojka. To achieve a reduction of the interference between the running kernels and to reduce the resulting execution jitter, we used a time-triggered execution on the GPU. To enable the synchronization, we assessed two synchronization mechanisms available on Tegra X2 platform: one based on zero-copy memory and one based on the globaltimer. We found that the NVIDIA profiler (nvprof) reconfigures the resolution of globaltimer from 1 s to 160 ns. With this resolution, we were able to reduce the execution time jitter of a tiled 2D convolution kernel from 6.47% to 0.15% while maintaining the same average execution time by use of a time-triggered GPU execution.