Functional Heterogeneous Processor Affinity Characterization to Big Data: Towards Machine Learning Approach

摘要

Heterogeneous processors based on functionally diverse devices (GPP, GPU, FPGA, ASIC and DSP) may provide affinity for different types of big data processing. In this paper, we developed affinity characterization of such functionally diverse device based heterogeneous systems. However, extracting such characterization is computationally intensive and rather slow for large set of nodes. We first developed general affinity mapping approach which showed 2 to 15 speed up and reduction of energy from 30-35%. We investigated the machine learning approach to the characterization by evaluating predictive performance using training data from disparate heterogeneous computing systems and achieve an improvement in affinity matching. We can extend this machine learning approach to develop comprehensive affinity between big data applications across several domains to heterogeneous computing using functionally diversified processors.