基于微服务架构FPGA云平台的并发请求调度机制

基于微服务架构的现场可编程门阵列（FPGA）云平台在被大规模推广后积累了许多用户。针对FPGA云平台存在大量用户并发请求的问题，建立一种基于优先级调度的自定义参数响应指数计算模型。将5个请求关键影响因素作为自定义参数，采用层次分析法确定各参数权重，根据响应指数函数计算各请求的响应指数。在该模型的基础上，设计一种高并发请求调度（HCRS）算法，通过响应指数阈值对请求类别进行划分，使得高优先级请求优先得到处理，次优先级请求加入先进先出队列等待，低优先级请求暂时挂起，从而缩短请求响应时间以及请求响应延时，缓解由高并发请求带来的硬件节点资源分配压力。在真实运营的FPGA云平台中实现该算法并在实际环境中进行测试，结果表明，在并发请求量相同时，与先来先服务调度算法相比，HCRS算法的平均响应延时降低29 074 ms，平均请求响应时间缩短12 605 ms，其在提升系统吞吐量与并发度的同时可以有效优化硬件节点资源利用率。

Concurrent Request Scheduling Mechanism for FPGA Cloud Platform Based on Microservice Architecture

The Field Programmable Gate Array(FPGA) cloud platform based on microservices architecture has accumulated many users following its large-scale rollout.It has a problem with a large number of concurrent user requests.To address this problem, this paper establishes a calculation model of user-defined parameter response index based on priority scheduling, determines five request key influencing factors as custom parameters, uses the Analytic Hierarchy Process(AHP) to determine the weight of each parameter, and calculates the response index of each request according to the response index function.Then, a High Concurrent Request Scheduling(HCRS) algorithm is designed based on this model.This algorithm classifies requests using a response index threshold.Therefore, the highest priority requests are processed first, secondary priority requests are added to the first-in-first-out queue to wait, and low priority requests are temporarily suspended.This method shortens the request response time and latency, easing the hardware node resource allocation pressure caused by high concurrent requests.Experiments compared the performance of the HCRS algorithm with the First Come, First Serve(FCFS) algorithm on the FPGA cloud platform in a real scenario.The results show that average response latency of the HCRS algorithm decreased by 29 074 ms, and the average request response time decreased by 12 605 ms.The proposed contribution improves throughput, system concurrency, and utilization of hardware node resources.