Research On Virtual Machine VCPU Scheduling For Performance Optimization

In data centers,the computer resource utilization is very low.To reduce the cost,virtual-ization has become the focus of attention.Virtualization technology abstracts and shares the underlying physical resources in the form of resource pools.To achieve different scheduling goals,these shared resources are properly allocated to different virtual machines(VM)or containers.Resource scheduling has two levels:one is resource allocation and management at the global level;the other one is task scheduling at the node level.At the node level,mul-tiple workloads share resources of the same physical node.In order to optimize and isolate performance,it usually needs to make the appropriate task scheduling based on the charac-teristics of the workloads and the underlying physical resources.However,the introduction of the virtualization layer brings many problems and challenges for the task scheduling.Based on a comprehensive analysis of the research status of resource scheduling in data centers,this dissertation focuses on the research of task scheduling at the node level.The main research contents are as follow:Traditional VCPU(Virtual CPU)schedulers only focus on fairly sharing CPU resources among VMs,while ignoring the system performance,which seriously reduces I/O perfor-mance.Although recent research significantly improves I/O performance,they degrade the performance of computational tasks by shortening time slices and reducing cache efficiency.To address the problems,this dissertation presents a multicore periodical preemption(MPP)VCPU scheduling scheme with three optimization techniques:(1)periodically coalescing and handling I/O events to reduce the scheduling latency,which guarantees I/O performance;(2)taking advantage of multicore environments and centrally handling I/O events on differ-ent cores in a Round-Robin manner to lengthen time slices,which improves the performance of computational tasks;(3)using a dedicated priority for I/O event handling to keep the CPU fairness.The experimental results demonstrate that our scheduling scheme efficiently lengthens time slices and improves the performance of computational tasks,and achieves the same I/O performance as the existing approaches optimized for I/O.Multiprocessor VMs encounter synchronization problems such as lock holder preemp-tion(LHP)and lock waiter preemption(LWP),resulting in a significant degradation of sys-tem performance.Recent research that addresses this issue has some shortcomings.To address these problems,this dissertation proposes a lock-aware(LA)VCPU scheduling scheme.Our approach provides a technique for detecting the lock state of current VCPU in the VMM(Virtual Machine Monitor)layer,without modifying guest kernel.Based on this technique,the proposed approach gives preempted lock holders and waiters multiple continuously extra scheduling chances to release locks.Our scheduling scheme assures that the de-scheduled VCPUs are not lock holders or waiters with large probability.The ex-perimental results demonstrate that our scheduling scheme fundamentally eliminates lock holder preemptions and lock waiter preemptions,and guarantees total system performance undiminished as the number of VMs increases,while keeping the scheduling fairness.In the multicore virtualization environments,the existing scheduling approaches are not friendly to real-time VMs,and the existing research on real-time task scheduling for multicore environments is inadequate.To address the problem,this dissertation proposes a load-balanced real-time VCPU scheduling scheme(Load Balanced P-EDF,LBP-EDF).The proposed scheduling scheme adopts a partitioned-based pattern to avoid system competition overhead caused by global-based approaches.Meanwhile,the proposed scheduling scheme provides a mechanism of load balancing between the runqueues.With this mechanism,our approach allocates the system CPU resources reasonably and improves CPU utilization,and ensures the schedulability of the task set.The experimental results demonstrate that LBP-EDF scheduling scheme guarantees the schedulability of the task set and improves the num-ber of schedulable tasks and CPU utilization.