۱- Introduction

۲- Background

۳- Motivation

۴- Design

۵- Evaluation

۶- Related work

۷- Conclusion

References

Abstract

Live migration of virtual machines (VMs) is an indispensable feature in cloud computing systems due to the use of load balancing, system maintenance, power management, etc. However, it is usually an intractable problem to migrate the VMs running memory write-intensive workloads, because they rapidly generate plenty of dirty pages, which need to be resent many times. Migration failures or compromises with large downtimes usually result in severe consequences such as the violation of service level agreements, abortion of TCP connections, etc. To solve this challenging problem, we first demonstrate a novel observation: during the pre-copy migrations, some workloads have a large portion of “fake dirty” pages, which are unnecessary and wasteful to be resent. After exploring how they are generated, we propose to leverage the secure hash method to avoid resending them. Besides, to guarantee the successful accomplishment of migration, we further propose the intelligent hybrid migration. Our scheme leverages heuristic and automatically switches from pre-copy to post-copy at the near-optimal moment to obtain a short post-copy duration, thus alleviating post-copy’s inherent weaknesses. Evaluations show that our proposed scheme gets a significant performance improvement for VM migration. The workloads which fail to be migrated with pre-copy now accomplish migration quickly, with a total migration time from 27 s to 98 s. Besides, the intelligent hybrid migration remarkably shortens the post-copy duration by an extent from 43% to 60%, compared with the traditional hybrid method.

Introduction

As one of the most important techniques in cloud computing systems, virtualization [1] has many superiorities, such as abstraction from heterogeneous hardware, security isolation, convenience for management, etc. A key characteristic of the virtualization technology is live virtual machine (VM) migration [2,3], which refers to moving a running VM from one physical host to another. Live VM migration is powerful in data centers due to the use of load balancing [4,5], online system maintenance [6], fault tolerance [7, 8] and power management [9], etc. Inside data centers, the source and target nodes for migration often share the same storage through SAN (Storage Area Network) or NAS (Network-Attached Storage). So only the content of VM memory and the states of the other virtual devices (for example, the virtual CPU state) need to be transferred. Pre-copy [2,3] is the prevailing approach for VM migration. In a pre-copy migration, the memory content is sent from the source to the target node in several sequential iterations. In the nth iteration, only the pages that are written dirty in the (n−۱)th iteration are transferred. Thus, pre-copy tries to obtain the goal of a short downtime, during which the VM is temporarily suspended. However, it is usually an intractable problem to migrate VMs running memory write-intensive workloads [10]. These workloads often persistently occupy a large fraction of system memory, while writing the memory pages dirty very fast. This leads to the rapid generation of plenty of dirty pages, which need to be resent many times during pre-copy. So the total data transferred (TDT) is huge and the total migration time (TMT) is prolonged.