Multi-machine Memory Optimization Technology Research And Virtualization Environments

With the development of the computer science and the network technology, system virtualization is widely used in data centers, for the purpose of consolidating the existing hardware resource and maximizing the utilization by building virtual machine clusters on the data centers. The key point lies in deploying more virtual machines on a single server with no harm to the performance of the system. However, the lack of memory becomes the main bottleneck to achieve the goal. So it becomes critical to find some effective methods to minimize the memory demands of the virtual machines. Nowadays, most of the researches are focused on the memory optimization in a single monitor machine. As we all known, the potential of memory saving on a single server is very small because of the limitation of the total memory capacity. If we broaden our view, taking the total data center into consideration, it will be more prospective.Based on above reasons, the paper proposes a mechanism of memory optimization among multi-machines in the environment of virtualization. In this method, all the physical memory of different machines joint together to make a large dynamic memory pool (DMP) where all virtual machines run. Then the new memory optimization mechanism can run on the DMP. DMP is a key widget in the method, first, it provides memory service, such as memory allocation and memory free, to the upper layer; second, it provides interfaces for the users to manage the physical memory. And the most important function of DMP is memory address transform. When DMP receives a logical address from the virtual machine, it should transmit it into the physical memory address. If there is any exception, for example, page fault interrupt, occurred, DMP should rescue the exception and recover the operation. Modify the content-based memory sharing functions, search and share the memory pages with the same content in the global memory space to save more memory. At the same time, the memory access function should be improved to provide faster service. In this paper, two methods are used, multi-Markov chains prefetching model and active memory. The former method is to fetch the memory pages ahead which predicted by the multi-Markov chains prefetching model, while the latter one is used to relieve the calculating pressure of memory server. The content-based memory sharing method can gain some hints of which virtual machine shares more pages with another.The functions described above are implemented as prototype based on KVM, and the experiments proved that the mechanism can save more memory, and the implemented system is more effective.