应用感知的数据中心网络多租户共享方法  

Application-Aware Network Sharing for Multi-Tenant Data Center

作　　者：李妍[1] 郭得科[1] 曹晓丰 陈洪辉[1] LI Yan;GUO De-Ke;CAO Xiao-Feng;CHEN Hong-Hui(Science and Technology on Information Systems Engineering Laboratory,National University of Defense Technology,Changsha 410073)

机构地区：[1]国防科技大学信息系统工程重点实验室,长沙410073

出　　处：《计算机学报》2021年第7期1363-1377,共15页Chinese Journal of Computers

基　　金：国家重点研发计划(2018YFE0207600);国家自然科学基金联合基金重点项目(U19B2024)资助.

摘　　要：在多租户数据中心,来自不同租户的应用程序共享并竞争使用网络资源.网络共享策略会对应用程序端到端的性能(如作业完成时间、吞吐量等)产生直接的影响.为了衡量租户应用程序的整体数据传输速率,本文引入进度(Progress)的指标.该指标被定义为租户应用程序在所有链路上经需求标准化后的最小带宽分配量,反映的是租户能够完成其数据传输的最慢速率.通过最大程度地提高租户进度,可以优化上层应用程序的执行时间等性能.先前的大多数工作都集中在实现网络共享的公平性、可预测性和效率之间的权衡,却忽略了提高租户的长期进度.本文观察发现应用程序放置于租户所租赁的不同虚拟机上会形成不同的带宽需求分布,进而影响后续带宽分配所能够获得的最优进度.通过理论分析我们证明了获得所有租户最优进度的关键在于最小化网络瓶颈链路上的带宽需求.基于此,本文提出应用感知的网络多租户共享方法,通过联合优化任务放置和带宽分配的过程,该方法最大化所有租户的进度,并在优势资源公平性限制下最大化网络利用率.实验结果证明,与目前的最新带宽分配方法相比,本文将租户整体进度提高了85.6%~107.7%,网络链路利用率提高了71.2%~112.4%.In the multi-tenant shared cloud,tenants usually deploy their applications on virtual machines(VMs)rent from the cloud.Such applications generate varying amounts of traffic between their VM pairs.Applications of different tenants compete for the shared network,resulting in significant unpredictable performance which not only makes tenant costs incalculable but also causes revenue loss for cloud service providers.Sharing network among tenants is challenging because the bandwidth obtained by tenants depends on a variety of complex factors,such as system load,VM placement,and the oversubscription ratio of the data center network.There is a broad consensus is that network sharing mechanisms should be fair,predictable,and efficient.Specifically,tenants expect to guarantee the minimum bandwidth to achieve performance predictability,while cloud providers pursue high network utilization and strategy-proofness.Research shows that there are strong trade-offs among these three goals.Most previous work has focused on achieving tradeoffs while ignoring the long-term goal of improving application performance.In fact,network allocation decisions have a significant impact on the end-to-end performance(e.g.,job completion time,throughput,etc.)of applications.To measure the overall data transfer rate of a tenant,we introduce the metric progress,which is defined as the minimum demand-normalized bandwidth allocation on all network links.The progress is an essential metric to indicate how fast a tenant can complete data transfer.By maximizing the tenant’s progress,application performance such as execution time can be optimized.Currently,much work has proved that given the bandwidth demand vectors of all tenants,the state-of-the-art bandwidth allocation schemes i.e.,DRF and HUG,can obtain the optimal progress.However,we found that by changing the bandwidth demand distribution,the tenant’s progress and isolation guarantee can be further improved.The insight is to tune and optimize the placement of tenants’applications to cause th

关 键 词：网络共享 带宽分配 多租户 应用感知 任务放置 进度 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]