Method of QoS evaluation of FPGA as a service

The subject of study in this article is the evaluation of the performance issues of cloud services implemented using FPGA technology. The goal is to improve the performance of cloud services built on top of multiple FPGA platforms known as FPGA-as-a-Service (FaaS). Task: to analyze the delays in communications between host computer and FPGA; propose the steps of development to reduce the delay and perform the evaluation of the response time for the FPGA-based accelerator depending on number of involved cards; consider the reliability aspect of such systems implemented using programmable logic. According to the tasks, the following results were obtained. The FPGA-as-a-Service where FPGA resources are provided through a set of hardware/software toolset is considered. The usage of queueing theory for cloud-based services is analyzed. The contribution of the parts of FPGA-as-a-Service to the final delay of the service is discussed. The process of modeling of work the services based on FPGA accelerator cards with use of Jackson's network is analyzed in detail. The model of the delays of FaaS that considers the parameters of accelerator FPGA cards is offered. The formula of the total response time of the service combined based on the response of the components of is obtained. The proposed steps of reduce data processing delays include increase the size of data blocks for processing in FPGA by each kernel, change the communication model with kernel from sequential to pipelined, following timing closure technique and use more FPGA accelerator cards in parallel to divide the enquiring delay. Based on the proposed model the evaluation of response time of FaaS was done. The advantage of the use of many FPGAs in parallel for same data processing task instead of implementation of requests thread for each accelerator card is shown. Conclusions. The main contribution of this study is a step forward to the modeling of FPGA-based services that can be used for FPGA-based artificial intelligence (AI) applications. It helps to improve the performance of the system by means of reducing the delays at different stages of requests processing. Another side of this result is the reliability aspect that is based on modified manner of service operation in case of use the proposed steps of system optimization. It helps to improve the processing of requests to FaaS. The proposed method is the next step after prototyping of such systems because it helps to turn the FaaS from the object for development to the tool for deployment of new technologies like AI applications.