基于跨虚拟机零下陷通信的加速器虚拟化框架  

Accelerator Virtualization Framework Based on Inter-VM Exitless Communication

作　　者：李鼎基 糜泽羽 吴保东 陈逊 赵永望 丁佐华 陈海波[1] LI Ding-Ji;MI Ze-Yu;WU Bao-Dong;CHEN Xun;ZHAO Yong-Wang;DING Zuo-Hua;CHEN Hai-Bo(School of Software,Shanghai Jiaotong University,Shanghai 200240,China;Sensetime,Beijing 100080,China;School of Cyber Science and Technology,Zhejiang University,Hangzhou 310007,China;School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]上海交通大学软件学院,上海200240 [2]北京市商汤科技开发有限公司,北京100080 [3]浙江大学网络空间安全学院,浙江杭州310007 [4]浙江理工大学信息学院,浙江杭州310018

出　　处：《软件学报》2020年第10期3019-3037,共19页Journal of Software

基　　金：广东省重点领域研发计划(2020B010164003);国家杰出青年科学基金(61925206);上海市科委高技术支持计划(19511121100)。

摘　　要：人工智能技术的长足发展对于云计算的算力提出了更高的要求,云服务提供商在数据中心内添置了拥有大量并行计算单元的加速器,这些加速器需要与已有的虚拟化平台相结合以进行计算资源的划分.当前主流的加速器虚拟化方案是通过PCI透传的方式,但是该方式不支持细粒度的资源划分;部分特定型号的加速器还支持了时分复用的方案,通过硬件与虚拟机监视器配合划分计算资源和时间片,但是该方案可移植性差,对于任何新型加速器的适配都要重新开发,固定的资源划分策略也导致可扩展性有限;另有基于API转发的方案,通过分离式驱动的模式将虚拟机的请求转发给后端驱动处理,而转发通信的过程中存在着性能瓶颈.提出了Wormhole,一种基于C/S架构的、支持跨虚拟机快速代理执行的加速器虚拟化框架,旨在为上层用户提供高效、透明的加速器API转发虚拟化的同时保障多用户间的强隔离性.该框架利用硬件虚拟化技术,允许CPU控制流在虚拟机间快速切换而不触发任何下陷,大幅降低了虚拟机间通信带来的虚拟化性能开销.实验结果表明,Wormhole的原型系统相较于具有代表性的开源虚拟化方案GvirtuS,在经典模型的训练测试中能够有高达5倍的性能提升.The increasing deployment of artificial intelligence has placed unprecedent requirements on the computing power of cloud computing.Cloud service providers have integrated accelerators with massive parallel computing units in the data center.These accelerators need to be combined with existing virtualization platforms to partition the computing resources.The current mainstream accelerator virtualization solution is through the PCI passthrough approach,which however does not support fine-grained resource provisioning.Some manufacturers also start to provide time-sliced multiplexing schemes,and use drivers to cooperate with specific hardware to divide resources and time slices to different virtual machines,which unfortunately suffer from poor portability and flexibility.One alternative another but promising approach is based on API forwarding,which forwards the virtual machine's request to the back-end driver for processing through a separate driver model.Yet,the communication due to API forwarding can easily become the performance bottleneck.This study proposes Wormhole,an accelerator virtualization framework based on the C/S architecture that supports rapid delegated execution across virtual machines.It aims to provide upper-level users with an efficient and transparent way to accelerate accelerator virtualization with API forwarding while ensuring strong isolation between multiple users.By leveraging hardware virtualization feature,the framework minimizes performance degradation through exitless cross-VM control flow switch.Experimental results show that Wormhole’s prototype system can achieve up to 5 times performance improvement over the classic open-source virtualization solution such as GVirtuS in the training test of the classic model.

关 键 词：虚拟化 加速器 人工智能 代理执行 虚拟机间通信 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]