面向云计算性能保证的多租户数据中心网络带宽分配与最优性价比计价体系研究

The essence of cloud computing is to allow multiple tenants to share computing, storage, and bandwidth resources. Most resources, such as CPU cycles, memory, and storage space, are shared among multiple tenants using virtualization: each application is hosted by one or multiple virtual machines (VMs), which are priced according to the amount of resources allocated and the amount of time used. Yet, network bandwidth between VMs, noticeably as a critical resource in datacenter networks, is largely used in a best-effort fashion in current-generation cloud datacenters. The amount of network bandwidth allocated to an application is affected by available bandwidth at the bottleneck along the network path between two VMs, subject to cross traffic from other applications. Due to the uncertainty in sharing the datacenter network in such a best-effort manner, a tenant cannot predict the finishing times of its tasks. In order to attract tenants to host their mission-critical applications in the cloud, this research proposal aims to achieve two fundamental objectives in cloud computing: the predictability of tenants' performance and the payment based on the actual bandwidth usage. To offer better performance predictability, tenants should be given the ability and flexibility to express their desired services. Specifically, we define a representative set of aggregation network bandwidth allocation classes: bandwidth guarantees, high priority provisioning, and fine-grained best-effort sharing, in decreasing order of predictability. To achieve better fairness among tenants in these aggregation classes, we aim to design a differentiated pricing framework to price bandwidth in a datacenter network, following a fundamental principle that we designed as "cost dominant fairness," in that a tenant will pay proportionally more if its applications have enjoyed better performance using more bandwidth resources. By applying advanced software-defined networking (SDN) techniques in an experimental datacenter testbed, we will validate the effectiveness of the proposed network bandwidth allocation and new pricing framework, based on realistic cloud application workloads. We believe that this research can achieve optimal utilization of cloud resources and economic cost-performance tradeoff of tenants.

当前云计算数据中心面向多租户的"尽力式"网络分享与单一计价模式，无法提供租户迫切关注的应用性能保证以及依据实际网络带宽消耗的最优性价比计价体系。为了吸引租户在云平台托管更关键的应用任务并完善租户与运营商双赢的公平云计算市场，本项目将提出以租户为主体的数据中心网络带宽权重分配机制及差异化计价新方法，实现不同租户各自分享的网络资源和实际性能与费用的适配性，以及租户间的相对公平性。包括：构建按照应用性能需求级别从高到低的多样化网络带宽分配服务聚类（包括硬性带宽保证、高优先级带宽供应、细粒度尽力带宽分配），提供灵活适配的网络资源需求表达与分配机制；设计基于博弈论和"成本优势型公平"原理的数据中心网络优化计价算法与控制协议，保证各租户按实际消耗带宽大小和性能体验优劣，来决定支付费用高低；实现基于软件定义的网络新技术的数据中心实验平台，以真实云应用验证面向服务质量保证的带宽分配公平性及计价合理性目标。

当前云计算数据中心面向多租户的“尽力式”网络分享模式无法提供租户迫切关注的应用性能保证（或可预测性）以及依据实际网络带宽消耗的最优性价比。为了利于租户在云平台托管更为关键的应用任务并完善租户与云运营商双赢的公平云计算市场，本项目研究了以租户为主体的数据中心网络带宽权重分配机制，以及适应于多样化应用及其实际带宽需求的差异化计价基础，以促进不同租户各自分享的网络资源和实际性能体验与所付费用的适配性，以及租户间的相对公平性。本项目经过四年潜心深入的研究，研究工作进展顺利，完成了预期的研究目标，取得了丰富的研究成果：共发表学术论文13篇，其中IEEE/ACM Transactions论文5篇，包括IEEE/ACM Transactions on Networking、IEEE Transactions on Computers、IEEE Transactions on Big Data、以及横跨IEEE所属各学会或专业组所有领域的顶级刊物（CCF“前沿、交叉与综合领域”排名第1位的A类刊物）《Proceedings of the IEEE》（2014年102卷第1期）封面主题论文（于2015年入选ESI高被引论文），该刊物自1913年创刊至今有100多年历史，在领域内具有重要影响力（影响因子9.237）；其他SCI期刊论文2篇；CCF A类顶级国际学术会议INFOCOM论文1篇、CCF B类重要国际学术会议论文2篇和短文1篇、网络系统领域的新兴国际学术会议APNet 2017论文1篇；另外，受邀在国家自然科学基金委员会主办的综合指导性学术期刊《Science Foundation in China（“中国科学基金”英文版）》发表了科普性综述论文1篇。共申请国家发明专利4项，其中已获批国家发明专利2项、在审国家发明专利2项。项目组2人获湖北省优秀硕士学位论文奖、1人获湖北省优秀博士学位论文奖，项目负责人于2017年入选了国家优青，并获华中科技大学十大“科技新星”、“科技创新优秀指导教师”荣誉称号。