Workload Proxies and Their Optimization for Healthcare Analytics

Abstract

Workload analysis plays important role in understanding the problems in the design and development of server systems. This will help the architect to make predictions of the system behavior of the upcoming system or SoC. Today market is moving towards the workloads related to healthcare image processing because healthcare is on high priority sector and people expect highest level of care and services regard- less of cost. This work primarily concentrates on the creation, characterization and optimization of workload proxies like U-Net, Xception and DenseNet for healthcare image processing on Intel Xeon Server Platform. In this thesis, for image segmen- tation U-Net topology is used and for image classification Xception and DenseNet topologies are used as source of the workload. With the help of few Intel internal tools and silicon data from the present generation platform, study is made to under- stand the behavior of the workload in present generation platform. We are able to measure the performance of different platforms for different workloads, to identify the bottlenecks in the performance of existing platforms or existing software, to find the reason for this bottleneck and predict the possible solution. So that this anal- ysis will be used to propose SoC architecture features and important optimizations needed to support a new class of workloads efficiently in next generation platforms or SoCs. Open Visual Inferencing and Neural Network Optimization (OpenVINO) and Multi-instance implementation provided significant performance boost. Central Processing Unit (CPU) and Graphics Processing Unit (GPU) performance com- parison results are also discussed in the last section. These workload proxies and performance numbers will be provided as proof of concept to core architects.