Optimize Virtual Machine On Multi-core Processor

The development of Internet has brought people the new compute methods, and cloud computing is one of them. Cloud computing consolidates the computing resources to provide virtualized computing environment to users. These virtualized computing resources are delivered quickly, managed flexibly and can be reused. Further more, the consolidation maximizes the utility of hardware to avoid energy wasting caused by the idling machines. As one of the key technologies of cloud computing, virtualization technology makes flexible resource management possible in cloud computing. It can serve several virtual machine (VM) instances in a single physical server.VM management tools are important parts in virtualization. With the help of VM management tools, VM can be saved, restored and migrated to other servers. With the development of multi-core processor, the ability of parallel computing improves a lot. But nowadays VM management tools are usually implemented in single thread model, which doesn’t fully utilize the multi-core processor. For the size of memory and persistent storage grows, cloud computing is adopting larger VM instance to meet users’requests, which leads to longer VM management time. If we parallelize the process of VM management tools, we can shorten the time of VM management on multi-core processor platform. As a result, cloud computing would confirm to provide flexible service.Multi-core processor also makes memory bandwidth lower for each processor core, and NUMA is adopted widely to solve this problem. However, virtualization technology hides the hardware detail when providing resource abstraction, which maks VM unaware of the real hardware configureation, such as NUMA topology. It causes memory management and task scheduling harder in hypervisor.We propose two methods to solve the problem above:1) Parallel VM management method, which parallelize the steps of VM management tool. It also adopts multi-thread model to map guest VM’s memory space, which takes the most time in the whole procedure.2) Dynamic NUMA method, which tells VM its NUMA topology by virtual firmware. VM can also dynamically reconstruct its NUMA topology when hypervisor schedule its VCPU to other NUMA node. In this way, we confirm that the NUMA toplogy in VM is accurate, which leads to less cross NUMA node memory access.To proof these two methods really improve the performance of VM management tools and VM itself, we implement these two methods on Xen virtualization platform and perform some comparison tests. The evaluation results are listed bellow:1) The parallel VM management method is faster than the default serialized one, by fully utilizing multi-core processor. The execution time is shorton to 33% of origin one in some test cases.2) The dynamic NUMA method makes VM run faster than the default configureation on NUMA hardware. And the performance of mmap test improves nearly 100%.