Super Multiplicative VRS DEA (SMVDEA) input-oriented model with raw measurements results

Super Multiplicative VRS DEA input-oriented model results obtained with 660 DMUs using apache bench tool for evaluating video stream files on two main webservers from the market, that are Apache2 vs Nginx, varying 15 TCP congestion control algorithms on Xen hypervisor.

Abstract:
The prediction of the best set of tools for delivering network services is always a hot topic of research, mainly on the Internet where the population changes its interest in the type of services to be consumed rapidly. Moreover, randomness is a challenge for all kinds of prediction, even though specifically on computer networks in the 90´s decade the self-similarity concept on computer networks brought the memory over time on spatiotemporal data. The self-similarity's memory and fractal dimension are important indexes to see the difference in performance and stability among a set of time series data over time. Recently, it was discovered on virtual networks that change the set of tools or virtual network hypervisors for delivering services there is a distinct fractal behaviour on network services providers. Then, the choice of the best virtual network hypervisor is mandatory to deliver the best way to provision virtual network services along the time. By the way, the researchers discovered along with the history a series of analytic ways of forecasting services based on time series, one of them is data envelopment analysis (DEA). DEA are nonparametric multicriteria decision-making models that use decision variables to make an accurate judgment of decision-making units (DMU). The main contribution of the paper is to devise a new super-efficiency multiplicative DEA model for the prediction of the best DMU varying the TCP congestion control algorithms and virtual network settings as vCPU (1, 2, 3, and 4) and vRAM (1,2, and 4) with network and fractal indexes as decision variables. The experiments were conducted on Xen type-I hypervisor to acquire the web traffic by using the apache-bench benchmarking tool for obtaining each decision variable from DMUs for DEA ranking.

本数据集涵盖基于660个决策单元（Decision Making Unit，DMU）、采用投入导向型超乘子可变规模报酬数据包络分析模型的实验结果。实验依托市场上两款主流Web服务器Apache2与Nginx，使用Apache Bench基准测试工具开展视频流文件性能评估，在Xen虚拟机管理程序环境下配置15种TCP拥塞控制算法以采集相关数据。

摘要：
面向网络服务交付的最优工具集预测始终是研究热点，在互联网场景中尤为显著——用户对所需消费的服务类型的兴趣正快速更迭。此外，随机性是各类预测任务共同面临的挑战；早在20世纪90年代，计算机网络领域提出的自相似性概念，便揭示了时空数据随时间推移所具备的记忆性特征。自相似性的记忆性与分形维数，是衡量一组时间序列数据随时间推移所展现出的性能与稳定性差异的核心指标。
近期研究发现，在虚拟网络场景中，若变更服务交付所使用的工具集或虚拟网络管理程序，网络服务提供商的网络行为将呈现出显著不同的分形特征。因此，为实现虚拟网络服务的长期最优部署，选择最优的虚拟网络管理程序是必要前提。
回顾研究历程，学者们已提出一系列基于时间序列的服务预测分析方法，数据包络分析（Data Envelopment Analysis，DEA）便是其中之一。DEA属于非参数多准则决策模型，通过决策变量对决策单元（DMU）进行精准评判。
本研究的核心贡献在于，提出一种全新的超效率乘子型DEA模型，用于预测最优决策单元；该模型以TCP拥塞控制算法、虚拟网络配置（如虚拟CPU（vCPU）数量为1、2、3、4，虚拟内存（vRAM）容量为1、2、4）以及网络与分形指标作为决策变量，充分考量了TCP拥塞控制算法与虚拟网络设置的变化情况。
本实验基于Xen Type-1虚拟机管理程序开展，通过Apache Bench基准测试工具采集网络流量，进而从各决策单元中获取DEA排序所需的各项决策变量。