EVALUATION OF OPENMP OPTIMIZATION IN HETEROGENEOUS COMPUTING MODE BY CODE OFFLOADING ON INTEL XEON PHI CO-PROCESSOR

Abstract

As the computing needs are increasing, the utilization of compute powers of multi-processors and co-processors together is an active area of research. This paradigm is known as Heterogeneous computing. With the increasing data sizes and complexity of algorithms, and dead lock reached in processor clock frequency due to power constraints, multi core and many core CPUs and GPUs have been used for parallel computing. This has become an inevitable approach for high volume data processing such as image processing. In the process of evaluation the authors had earlier tested on the Intel Quad Core i7 (8 threads) processor and dual Intel Xeon 12 core (48 threads) CPUs by optimizing an image processing code on K-Means clustering in multispectral feature space using remote sensing data[1]. The maximum speedup 5x is achieved on Intel i7 core CPU and speedup of 13x is achieved on Intel Xeon CPU by invoking dynamic scheduling when number of threads deployed are large. In continuation of the earlier studies, the present study analyses the Intel Xeon phi coprocessor 7120P(device) HPC accelerator performance with processor base frequency of 1.24 GHz along with OpenMP Parallel computing model. It is observed that the offloading will not give best result with small data size. To get the full benefits of offloading on Intel Xeon phi coprocessor, computation offloading with OpenMP utilizing both processor and coprocessor gains accelerations and increases the performance if communication overhead is less than the computation times which is highly application dependent.