Prediction of the best alternative for cloud network services delivering by using fractals variables and super-efficiency DEA models for time-series data comparison

This dataset is based on active measurements obeying RFC 2544 and RFC 6815, forming 150 tunneled virtual networks to be compared. The virtual networks were assembled using the VXLAN protocol in an IaaS within DevStack (ALL-IN-ONE) running on a VM in VirtualBox on an Ubuntu 20.4 operating system of a computer with an Intel i7 7500U processor, 2.9GHz clock, and 16GB of RAM. OpenStack-Neutron with virtual switch SDN (OpenvSwitch) was used to interconnect the VMs under analysis. The benchmarking tools iperf, VBoxManage, and lscpu (to capture the temperature of each core) were used for active measurements. In the experiments, two Guest VMs were utilized, with one serving as a traffic generator (TG) and the other as a device under test (DUT), but the DUT runs within a Docker container acting as a server. In the DUT, we varied the number of vCPUs (1,2,4), the amount of vRAM (512MB, 1GB, 2GB, 4GB, and 8GB), and the fifteen Linux TCP congestion control algorithms (BIC, BBR, CDG, CUBIC, DCTCP, ILLINOIS, HYBLA, HTCP, LP, NV, VEGAS, VENO, SCALABLE, WESTWOOD, and YEAH) making up the 150 decision-making units (DMU). The virtual networks (DMU) were ranked using the super-efficiency model of data envelopment analysis (DEA) with variable return to scale and input-oriented. The decision variables listed for ranking were: a) inputs – X1) TCP bandwidth fractal dimension; X2) core0 fractal dimension, X3) core1 fractal dimension, X4) core0 temperature average, and X5) core1 temperature average, and b) outputs– Y1) core 0 hurst, Y2) core 1 Hurst, Y3) TCP bandwidth average, and Y4) TCP bandwidth Hurst. All data per DMU were separated by a folder containing their respective time series. Thus, we conclude that the correct DMU was chosen to provide optimal virtual network services over time. All scripts used to extract decision variables are available in this dataset. Calculating fractal dimensions was madogram and the Hurst parameter was calculated using R/S.本数据集基于遵循RFC 2544与RFC 6815标准的主动测量方法，构建了150个隧道化虚拟网络用于对比分析。本次实验的虚拟网络搭建于运行于Ubuntu 20.4操作系统的DevStack（ALL-IN-ONE）架构的基础设施即服务（IaaS, Infrastructure as a Service）环境中，该宿主机搭载Intel i7 7500U处理器（主频2.9GHz）与16GB内存，其上的VirtualBox虚拟机内部署了该DevStack环境。虚拟网络通过虚拟扩展局域网（VXLAN, Virtual Extensible LAN）协议组装，并通过搭载软件定义网络（SDN, Software Defined Network）虚拟交换机OpenvSwitch的OpenStack-Neutron组件实现待测虚拟机（VM, Virtual Machine）的互联。本次主动测量采用iperf、VBoxManage以及用于采集各CPU核心温度的lscpu工具开展实验。实验中共使用两台访客虚拟机（Guest VM），其中一台作为流量生成器（TG, Traffic Generator），另一台作为待测设备（DUT, Device Under Test），且待测设备运行于充当服务器的Docker容器内。针对待测设备，我们调整了其虚拟CPU（vCPU）数量（1、2、4）、虚拟内存（vRAM）容量（512MB、1GB、2GB、4GB及8GB），以及15种Linux TCP拥塞控制算法（BIC、BBR、CDG、CUBIC、DCTCP、ILLINOIS、HYBLA、HTCP、LP、NV、VEGAS、VENO、SCALABLE、WESTWOOD、YEAH），由此形成150个决策单元（DMU, Decision-Making Unit）。本研究采用基于投入导向、可变规模报酬的数据包络分析（DEA, Data Envelopment Analysis）超效率模型对各虚拟网络（即决策单元）进行排序。本次排序所用的决策变量包括：a) 投入指标：X1）TCP带宽分形维数；X2）核心0分形维数；X3）核心1分形维数；X4）核心0平均温度；X5）核心1平均温度；b) 产出指标：Y1）核心0赫斯特指数；Y2）核心1赫斯特指数；Y3）TCP带宽平均值；Y4）TCP带宽赫斯特指数。每个决策单元的全部数据均存储于对应专属文件夹中，内含其对应的时间序列数据。据此可筛选出最优决策单元，以实现长期稳定的优质虚拟网络服务。本数据集公开了所有用于提取决策变量的脚本代码。本次实验中，分形维数通过madogram方法计算，赫斯特参数则采用R/S分析法求解。