Abstract

Cloud computing is the new trend in computing services and IT industry, this computing paradigm has numerous benefits to utilize IT infrastructure resources and reduce services cost. The key feature of cloud computing depends on mobility and scalability of the computing resources, by managing virtual machines. The virtualization decouples the software from the hardware and manages the software and hardware resources in an easy way without interruption of services. Live virtual machine migration is an essential tool for dynamic resource management in current data centers. Live virtual machine is defined as the process of moving a running virtual machine or application between different physical machines without disconnecting the client or application. Many techniques have been developed to achieve this goal based on several metrics (total migration time, downtime, size of data sent and application performance) that are used to measure the performance of live migration. These metrics measure the quality of the VM services that clients care about, because the main goal of clients is keeping the applications performance with minimum service interruption.
The pre-copy live VM migration is done in four phases: preparation, iterative migration, stop and copy, and resume and commitment. During the preparation phase, the source and destination physical servers are selected, the resources in destination physical server are reserved, and the critical VM is selected to be migrated. The cloud manager responsibility is to make all of these decisions. VM state migration takes place and memory state is transferred to the target node during iterative migration phase. Meanwhile, the migrated VM continues to execute and dirties its memory. In the stop and copy phase, VM virtual CPU is stopped and then the processor and network states are transferred to the destination host. Service downtime results from stopping VM execution and moving the VM CPU and network states. Finally in the resume and commitment phase, the migrated VM is resumed running in the destination physical host, the remaining memory pages are pulled by destination machine from the source machine. The source machine resources are released and eliminated.
In this thesis, pre-copy live VM migration using Distributed Shared Memory (DSM) computing model is proposed. The setup is built using two identical computation nodes to construct all the proposed environment services architecture namely the virtualization infrastructure (Xenserver6.2 hypervisor), the shared storage server (the network file system), and the DSM and High Performance Computing (HPC) cluster. The custom DSM framework is based on a low latency memory update named Grappa. Moreover, HPC cluster is used to parallelize the work load by using CPUs computation nodes. HPC cluster employs OPENMPI and MPI libraries to support parallelization and auto-parallelization. The DSM allows the cluster CPUs to access the same memory space pages resulting in less memory data updates, which reduces the amount of data transferred through the network.
The thesis proposed model achieves a good enhancement of the live VM migration metrics. Downtime is reduced by 50 % in the idle workload of Windows VM and 66.6% in case of Ubuntu Linux idle workload. In general, the proposed model not only reduces the downtime and the total amount of data sent, but also does not degrade other metrics like the total migration time and the applications performance.