基于通用多核处理器的5G应用并发场景行为分析

Analysis of 5G Workload’s Performance Based on General Multi-core Processor in Concurrent Scenario

由于5G通信场景具有大带宽、低延迟、海量流量和多样性等特征, 5G业务由传统基站承载转向数据中心部署已成为趋势.为给出5G应用在数据中心的部署建议,以开源OpenLTE作为其代表性benchmark进行分析,由于开源的OpenLTE性能很差,在性能分析时不能反映真实场景和行为特征,因此首先根据通用处理器的特点对其代码进行了优化,取得了2.5倍性能加速比;在此基础上结合处理器特征分析应用执行行为,其主要特点为:5G下行过程物理层处理流程是计算密集的,最高端口利用率90%,访存不密集,程序响应时间极短;最后结合通用处理器独特参数(多核、SMT和Turbo Boost等)分析应用在并发场景下的行为表现,并以提升数据中心资源率为目的给出部署建议, 5G应用的强实时性使其只能以独占机器方式运行,其内部并发体之间对共享缓存和访存带宽竞争小而对执行部件竞争激烈,可采用并发量不多于处理器核数方式部署,同时TurboBoost的影响不可忽视.

Due to the characteristics of large bandwidth, low latency, massive traffic, and diversity in 5G communication scenarios, it has become a trend for 5G services to shift from traditional base stations to data centers. To solve the deployment problem of 5G workloads, we use the open-source OpenLTE for performance analysis as its representative benchmark. Since open-source OpenLTE has poor performance, and it cannot reflect real scenarios and behavioral characteristics in performance analysis, we first optimize its codes according to the characteristics of general-purpose processors(GPPs) and achieve the speedup of 2.5x. On this basis, the workload execution behavior is analyzed considering the characteristics of GPPs, and the analysis shows that in the physical layer, the processing flow of the 5G downlink process is computationally intensive with a maximum port utilization rate of up to 90%, and memory access is not intensive with extremely short program response time. Finally, we analyze the workload behavior in concurrent scenarios in combination with the unique parameters of GPPs(multi-core, SMT, and turbo boost, etc.) and put forward deployment suggestions for improving data center resources. It is found that 5G workloads can only run in an exclusive machine mode on account of their strong real-time nature, and there is little competition among internal concurrent bodies for shared cache and memory access bandwidth but fierce competition for execution components. Thus, the concurrency value should not exceed the number of processor cores for deployment, and the impact of TurboBoost cannot be ignored.