Design And Implementation Of Performance And Cost Measurement Platform For Emerging Services In Geo-distributed Public Iaa S Clouds

Cloud service providers promote a new cloud product namely field programmable gate array(FPGA)infrastructure as a service(IaaS),which marks the new era of cloud IaaS.FPGA IaaS wraps virtualized CPU,memory,etc.and a specific FPGA board as a compute instance and rents it to users,e.g.,Amazon AWS F1.Data acceleration with FPGA instances can achieve extremely high throughput and low latency.Thus,FPGA instances are widely used to accelerate applications like machine learning and real-time video processing.Though FPGA IaaS services seem very promising,there still remain questions such as their real-world performance and best-fit scenarios.To reveal the real performance of FPGA IaaS,a performance and cost measurement platform for FPGA clouds is proposed.The measurement studies include analyzing the performance and cost differences of FPGA instances running computation-intensive and communication-intensive workloads,the impact of PCIe drives on inter-communication efficiency,as well as the performance degradation caused by virtualization and PCIe bandwidth competition.The measurements are deployed on Amazon AWS and Alibaba clouds covering three Xilinx and Intel FPGA instances.The experimental results show that:(1)It is ubiquitous that performance-cost anomaly exists on current FPGA IaaS platforms,i.e.,the performance does not increase linearly as the cost increases.(2)The inter-communication performance in FPGA clouds is tightly constrained by hardware drivers and virtualization technology.(3)The virtualized FPGA clouds are far from mature,e.g.,the performance of communication-intensive applications is terrible due to low PCIe bandwidth utilization.Our study not only guides tenants' selection of FPGA services,but also helps cloud service providers to improve the performance of FPGA clouds.