Research On Computing Productivity Optimization In Virtualization Environment

Virtualization makes it possible to let multiple operating systems run together on a single computer. In a virtualization environment, Virtual Machine Monitor (i.e., VMM) replaces the role of traditional operating system or act as part of host OS, running in the highest privilege level. VMM can create multiple isolated containers which are called virtual machines (i.e., VM). The co-existed operating systems are called guest OSes, run in VMs and be managed by VMM. Computing Productivity is an important metric of computer system, cosists of cost, performance, programmability, portability and robustness. As virtualization hasn't been matured very much, the productivity of virtualization environment is worse than traditional environment, mainly reveals at power comsumption and system performance. As a result, the emphasis of this research is how to optimize the productivity of virtualization environment by reducing power cost and improving performance. In detail, it consists of:VM-dominant power management, accurate time accounting of virtual machines, dynamic switching-frequency scaling and affinity-aware proportional share scheduling.For cost reduction, power management is a basic instrument of productivity optimization. Because costs are mainly from the continuous investments which grow with time, and the major part of continuous investments is electric energy (power). Virtualization is now spreading from the field of server to client. Virtual machines in client environment aren't equivalent, instead, there is a primary VM interacts with user and occupies most of CPU cycles. As a result, a VM-dominant power management solution is proposed. This method exposes the hardware power features, which are traditionally masked by VMM, to the primary VM. In this way, matured power management technology can be fully reused and VMM only need a few coordination logics. Experiment results show that VM-dominant power management can reduce power consumption by 8～22% than VMM-centric approaches.For performance improvement, the key of productivity optimization is CPU scheduling. To coordinate the two scheduling layers of Guest OS and VMM, an accurate time accounting method for hardware-assisted virtual machine, XenHVMAcct, is proposed. CPU time accounting provides the data source of performance measurement and process scheduling, and its accuracy will impact the reliability of measurement and the fairness of scheduling. CPU cycles are exclusively used by one OS in traditional environment, but shared by multiple VMs in virtualization environment. As a result, traditional time accounting method isn't appropriate for virtual machines. XenHVMAcct adds a SCHED_IN virtual interrupt in VMM, and there is a kernel module in Guest OS to handle it. XenHVMAcct significantly improves the accuracy of time accounting.Dynamic switching-frequency scaling is an optimization of CPU scheduling policy which is used to improve the performance of overcommitted VMs. Overcommitment is a usual phenomenon in virtualization, which means the number of virtual CPUs (VCPU) is more than that of physical CPUs (PCPU). In the widely-used Xen environment, if communication-intensive and I/O-intensive applications run in overcommitted VMs and default Credit scheduler is used, the performance will decrease seriously. The improved scheduling policy can monitor the behaviors of VCPUs by recording bus access events, and dynamically scale the frequency of context switching, so as to optimize the performance. Experiments show that the new policy can make overcommitted VMs obtain nearly the same performance as in undercommitted cases.Credit-APS scheduler is an affinity-aware fair scheduling policy based on Credit scheduler; it improves the fairness where there are affinity restrictions. CPU-affinity restricts a VCPU to run on a subset of all PCPUs, so it can obtain stronger performance isolation by reducing VCPU migration. Credit scheduler can achieve proportional share (i.e., fairness) in free-mapping case (no affinity restriction), but not so in restricted-mapping cases. Credit-APS scheduler is based on Credit algorithm, redefines the conception of weight, and is aware of CPU-affinity when allocating time slices to VCPUs, so achieves proportional share in all case.VM-dominant power management and accurate time accounting provide two basic mechanisms. On the other hand, dynamic switching-frequency scaling and affinity-aware scheduling are both improvements of CPU scheduling policies applied in different scenarios. These instruments effectively improve the computing productivity in virtualization environment.