XME: Semantic Based Memory Sharing Model for Xen Hypervisor

Abstract

Server virtualization has been identified as the key solution to the most of the issues related with modern computing environments. As a result virtualization technology has been becoming increasingly popular among the organizations. However in most virtualized environments virtual machines have been equipped with homogeneous operating systems, libraries and applications which create a significant redundancy in main memory at runtime. Dropping this memory redundancy is advantageous, since the reclaimed memory pages obtained by reducing private memory pages into sharable copies could be used to increase the memory reservation of each virtual machine or to start new virtual machines. Previous works have shown that Content-based page sharing(CBPS) could be used effectively in order to remove the memory redundancy occurs in virtualized environments and thus to reduce the memory footprint of virtual machines. However content-based page sharing has certain drawbacks such as requirement of computationally intensive memory scans and higher amount of of write-faults(CoW faults) occurred in sharable memory pages. As a solution some researchers have proposed to use semantics in page sharing. But the lack of runtime information within the virtual machine monitor regarding the guest operating systems, termed as semantic gap , has been the major obstacle to efficiently implement semantic-based page sharing approaches. However previous efforts have shown that this could be achieved, but at the expense of modified guest operating systems. Due to this fact, semantic-based page sharing approach has not been accepted by most of the manufacturers/developers of the virtual machine monitors. In this paper, the authors propose XMe, a novel semantic-based technique to overcome the issues with existing content-based page sharing approaches as well as semantic-based approaches. The novel approach is that instead of modifying the guest operating systems, we use the System map file in guest kernels to passively absorb the semantic knowledge required for page sharing. We design and implement our proposed solution as an extension to Xen hypervisor version 4.1.2- rc1 with the usage of Xen control libraries. The implementation is used to evaluate the effectiveness of our solution in terms of writefaults occurred in shared memory pages and performance overhead introduced to the system comparative to content-based page sharing approach. We have shown that our proposed approach significantly outperforms existing page sharing approaches in terms of reduction of both writefaults occurred in shared memory pages and performance overhead. Further our results have concluded that almost 99% of the memory pages in kernel code segments across VMs running identical linux kernels could be shared.