With the rise of High Performance Computing Systems usage for computationally intensive algorithms, it is necessary to examine these systems for the best configurations to run those algorithms efficiently. This thesis presents a study of the various setups and configurations of Intel Xeon Phi Knights Landing (KNL) for running Positron Emission Tomography (PET) scan image reconstruction algorithm, Maximum Likelihood Expectation Maximization (MLEM). It focuses on analyzing the effect of KNL specific hardware configurations, memory and cluster modes, as well as parallel programming models, such as Message Passing and Shared Memory, and techniques on performance. This yields 20 different setups to be compared to derive the best configuration to utilize KNL for running MLEM. The results show that, on one node, OpenMP implementation with affinity, implicit High Bandwidth Memory (HBM) usage model, and either All-to-all or Quadrant cluster modes outperform MPI-only and Math Kernel Library (MKL) implementation with respect to all KNL configurations. The assessment criteria are average runtime (performance), speedup (scalability), and bandwidth memory utilization. Nevertheless, it is worth to point out that implicit HBM usage model is obtained through linking Intel’s heap manager library, called memkind, for offloading to KNL’s HBM without any code changes; proving that it is easy to run code on KNL chipset.     «

With the rise of High Performance Computing Systems usage for computationally intensive algorithms, it is necessary to examine these systems for the best configurations to run those algorithms efficiently. This thesis presents a study of the various setups and configurations of Intel Xeon Phi Knights Landing (KNL) for running Positron Emission Tomography (PET) scan image reconstruction algorithm, Maximum Likelihood Expectation Maximization (MLEM). It focuses on analyzing the effect of KNL spec...     »

With the rise of High Performance Computing Systems usage for computationally intensive algorithms, it is necessary to examine these systems for the best configurations to run those algorithms efficiently. This thesis presents a study of the various setups and configurations of Intel Xeon Phi Knights Landing (KNL) for running Positron Emission Tomography (PET) scan image reconstruction algorithm, Maximum Likelihood Expectation Maximization (MLEM). It focuses on analyzing the effect of KNL specific hardware configurations, memory and cluster modes, as well as parallel programming models, such as Message Passing and Shared Memory, and techniques on performance. This yields 20 different setups to be compared to derive the best configuration to utilize KNL for running MLEM. The results show that, on one node, OpenMP implementation with affinity, implicit High Bandwidth Memory (HBM) usage model, and either All-to-all or Quadrant cluster modes outperform MPI-only and Math Kernel Library (MKL) implementation with respect to all KNL configurations. The assessment criteria are average runtime (performance), speedup (scalability), and bandwidth memory utilization. Nevertheless, it is worth to point out that implicit HBM usage model is obtained through linking Intel’s heap manager library, called memkind, for offloading to KNL’s HBM without any code changes; proving that it is easy to run code on KNL chipset.     «

With the rise of High Performance Computing Systems usage for computationally intensive algorithms, it is necessary to examine these systems for the best configurations to run those algorithms efficiently. This thesis presents a study of the various setups and configurations of Intel Xeon Phi Knights Landing (KNL) for running Positron Emission Tomography (PET) scan image reconstruction algorithm, Maximum Likelihood Expectation Maximization (MLEM). It focuses on analyzing the effect of KNL spec...     »