Optimizing virtual machine I/O performance in virtualized cloud by differentiated-frequency scheduling and functionality offloading

Many enterprises are increasingly moving their applications to private cloud environments or public cloud platforms. A key technology driving cloud computing is virtualization which can serve multiple VMs in one physical machine hence providing better management flexibility and significant savings in operational costs. However, one important consequence of virtualized hosts in the cloud is the negative impact it has on the I/O performance of the applications running in the VMs. In this dissertation, we demonstrate that the negative impact of virtualized hosts is mainly caused by two reasons. One is VM consolidation, the other one is virtualization device overhead. First, to alleviate the negative impact of VM consolidation on I/O performance, we introduce two solutions vSlicer and vTurbo. vSlicer enables more timely processing of I/O events by latency sensitive VMs (LSVMs), without violating the CPU share fairness among all CPU sharing VMs. vTurbo is a system that accelerates I/O processing for VMs by offloading I/O processing to a designated core, hence significantly improving the VMs network and disk I/O throughput. Second, we show that data movement in the cloud may incur tremendous overhead on different protection layers. Especially, when we directly move bigdata systems such as Hadoop to a virtualized cloud, we observe that device virtualization overhead affects I/O performance of the Hadoop distributed file system (HDFS). My developed work vRead , which enables ”direct” read to the disk image of HDFS datanode VM at the hypervisor layer, can avoid most of the virtualization associated overheads and hence improve the I/O performance of applications running in the VMs