Key Technologies Of Performance Optimization For Virtual Machine Live Migration

Live migration of virtual machines (VM) can seamlessly and transparently relocate a VM from physical machine to another machine. It is widely used for load balancing, power saving, fault tolerance and online maintenance, thus provides a significant benefit for administrators of data centers and clusters. VM live migration has became a key technology of virtualization, it provide significant advantages to construct flexible, scalable, and resource-efficiency large green datacenters. The research on virtual machine live migration is emerging in recent years, however, it have quickly attracted a lot of interest in research comites of virtualization technology. By now, there are still many issues that need to be further studied. First, there is still lack of performance prediction model for VM migration to guide migration decision making. Second, VM checkpointing, the technical support for VM migration, need to save a whole operation system, so the cost is relatively high. It needs optimizing techniques to develop a lightweight and transparent checkpointing mechanism. Third, the widely used pre-copying algorithm is not suit for all scenarios, for example, a memory intensive workload may lead to pre-copy inapplicable. At last, the performance of VM migration is mainly determined by the characteristic of applications and the selection of migration algorithms. There is still lack of a synthesized decision-making scheme to adaptively map different applications and different migration algorithms. This dissertation targets to the performance optimization of VM live migration, studies the key technologies of VM migration from four aspects.First, we propose the performance and energy prediction models for VM live migration. As migration costs may vary significantly for different workloads due to the variety of VM configurations and workload characteristics. Current study on VM placement mostly focus on the issue of trigger conditions (when) and destination selections (where), however, migration decision-making still lack model-guided methodologies to predict the migration cost. To make an optimal migration decision, we investigate design methodologies to minimize the migration overhead by quantitatively predicting the migration performance and energy cost. We thoroughly analyze the key parameters that affect the migration cost from theory to practice, and design two application-oblivious models for the cost prediction by using learning technique such as linear regression. To the best of our knowledge, this is the first kind of work to estimate VM live migration cost in terms of both performance and energy in a quantitative approach. The models should be one of criterias for migration decision making and foundation for algorithm design of VM placement. It can also guide the performance optimazation of VM live migration algorithm.Second, we propose free page evication, copy-on-write and increamental mechanism to optimize the performance of VM checkpointing, and implement a prototype for lightweight and live VM checkpointing, called VMckpt. Because the virtualization layer has little semantic knowledge about the operation system and the applications running atop, VM-layer checkpointing requires saving the entire operating system state rather than a single process. The overhead may render the approach impractical. To reduce the size of VM checkpoint, in this paper we propose a page eviction scheme and an incremental checkpointing mechanism to avoid saving unnecessary VM pages in the checkpoint. This scheme can significantly reduce the size of VM checkpoint, and thus decrease the time and space cost for saving and restoring a VM checkpoint. To keep the VM online transparently, we propose a live checkpointing mechanism by saving the memory image in a Copy-On-Write (COW) manner. The checkpointing downtime can be reduced to an imperceptible level.Third, we propose a novel algorithm of VM live migration based on VM live checkpointing and full-system logging/replay techiques, called CR/TR-Motion. Although previous memory-to-memory approaches demonstrate the effectiveness of live VM migration in local area networks (LAN), they are not applicable for all workloads and scenarios, especially for some memory-intensive workloads. Moreover, they would cause a long period of downtime in wide area network (WAN) environments. In this paper, we design and implement a novel approach, namely CR/TR-Motion, which adopts checkpointing/recovery and trace/replay technologies to provide fast, transparent VM migration for both LAN and WAN environments. In contrast to pre-copying algorithm, CR/TR-Motion synchronizes the migrated VM's state using execution trace logged on the source host instead of dirty memory pages, and the destination host replays the logs till the target VM gets a consistent state with the source VM. As the data volume of log transferred during synchronization is much less than dirty pages, CR/TR-Motion can greatly reduce the migration downtime and network bandwidth consumption compared to pre-copy algorithm. Moreover, because CR/TR-Motion algorithm only requires a litter network bandwidth, it is espercially applicable for VM migration in WAN environment.At last, we propose a decision-making model for adaptive migration algorithm seclection based on application characteristics. The performance of live VM migration is mainly determined by the characteristic of applications and the seclection of migration algorithms. Currently, there is still neck of a synthetic mechanism for migration decision making to adaptively map different migration algorithms and different types of workloads or application scenarioes. After a thorough analysis of features of different migration algorithms, we choose three typical migration algorithms in this model, which are pre-copying, post-copying and CR/TR-Motion. We first construct the performance prediction models for each migration algorithm, and then online sample and analyse the application features, which may directly guide the migration algorithm selection. For some instance, the requirement of migration performance is non-deterministic or vague, we design a mapping algorithm between different applications and migration algorithms based on fuzzy theory. A case study shows that the fuzzy synthetic decision making can always choose the optimal migration algorithm for a typical application.