Data transfer and reuse analysis tool for gpu-offloading using openmp

Abstract

In the high performance computing sector, researchers and application developers expend considerable effort to port their applications to GPU-based clusters in order to take advantage of the massive parallelism and energy efficiency of a GPU. Unfortunately porting or writing an application for accelerators, such as GPUs, requires extensive knowledge of the underlying architectures, the application/algorithm and the interfacing programming model, such as CUDA, HIP or OpenMP. Compared to native GPU programming models, OpenMP has a shorter learning curve, is portable and potentially also performance portable. To reduce the developer effort, OpenMP provides implicit data transfer between CPU and GPU. OpenMP users may control the duration of a data object’s allocation on the GPU via the use of target data regions, but they do not need to. Unfortunately, unless data mappings are explicitly provided by the user, compilers like Clang move all data accessed by a kernel to the GPU without considering its prior availability on the device. As a result, applications may spend a significant portion of their execution time on data transfer. Yet exploiting data reuse opportunities in an application has the potential to significantly reduce the overall execution time. In this paper we present a source-to-source tool that automatically identifies data in an OpenMP program which do not need to be transferred between CPU and GPU. The tool capitalizes on any data reuse opportunities to insert the pertinent, optimized OpenMP target data directives. Our experimental results show considerable reduction in the overall execution time of a set of micro-benchmarks and some benchmark applications from the Rodinia benchmark suite. To the best of our knowledge, no other tool optimizes OpenMP data mappings by identifying and exploiting data reuse opportunities between kernels.