ABSTRACT With the rapidly expanded biometric data collected by various sources for identi- cation and verication purposes, how to manage and process such Big Data draws great concern. On one hand, biometric applications normally involve comparing a huge amount of samples and templates, which has strict requirements on the computational capability of the underlying hardware platform. On the other hand, the number of cores and associated threads that hardware can support has increased greatly; an example is the newly released Intel Xeon Phi coprocessor. Hence, Big Data

CONTENTS 20.1 Introduction 394 20.2 Background 395

20.2.1 Intel Xeon Phi 395 20.2.2 Iris Matching Algorithm 396 20.2.3 OpenMP 397 20.2.4 Intel VTune Amplier 397

20.3 Experiments 398 20.3.1 Experiment Setup 398 20.3.2 Workload Characteristics 398 20.3.3 Impact of Dierent Anity 399 20.3.4 Optimal Number of reads 401 20.3.5 Vectorization 401

20.4 Conclusions 403 Acknowledgments 403 References 403

biometrics processing demands the execution of the applications at a higher parallelism level. Taking an iris matching algorithm as a case study, we implemented an open  multi-processing (OpenMP) version of the algorithm to examine its performance on the Intel Xeon Phi coprocessor. Our target is to evaluate our parallelization approach and the inuence from the optimal number of threads, the impact of thread-to-core anity, and the impact of the hardware vector engine.