| Graphic Process Unit(GPU)is the core for modern graphics computation.Due to its parallel nature GPU has potential to enhance the computation in a wide range.However,GPU has its limitations.For various reasons GPU may hang in applications,whilst full GPU virtualization does not support live migration like CPU in Xen.This has seviourly affected the efficiency of GPU applications.For the moment the state-of-the-art method to get round is to reset the GPU hardware via a specific mechanism provided by GPU vendors.The disadvantage of such reset operation is that it causes application unavailability in order to maintain stability of the operating system.This dissertation presents an efficient method to address the above limitations.In particular,we start from the gVirt method,and develop a novel live migration scheme to support high availability(HA)of full GPU virtualizsation.We define our scheme as High Availability gVirt(or gHA for short).The key achievement of the scheme is to build up a duplex structure under the full virtualized GPU environment.The proposed gHA backs up iteratively the whole Virtual Machine(VM),and migrates the VM to a new host.In circumstances when GPU hangs,the backup VM takes over to guarantee the high availability of the virtualized resources,and thus reduces the GPU limitations mentioned above.Not surprisingly,like all other application programs,cost and overhead occur when gHA scheme is applied.The many tests undertaken have shown that downtime of the VM backup is between 272 ms and 551 ms,only 48-327 ms more than the idle VM without GPU virtualization support,which is a quite good result compared to the 2-4s of the backup interval.For numerous GPU workload tests,our evaluation shows that different GPU workloads achieve 65.0-92.6% of gVirt performance,which is a good trade-off between performance and stability.Our solution occupies 80-180 Mbps bandwidth during execution,which is fairly minimal as the total bandwidth of 1Gbps between the two hosts. |
PDF Full Text Request