Efficient Pre-Copy Live Migration of Virtual Machines for High Performance Computing in Cloud Computing Environments

Pre-copy live migration is the most popular migration mechanism implemented in most hypervisors. However, it is not suitable for Virtual Machines (VMs) running computation-intensive and memory-intensive workloads because its operational behaviors depend on a configuration parameter, namely the maximum tolerable downtime. Although the default value of this parameter is good enough for normal web-based applications, it is too low for most HPC applications, which are computation-intensive and memoryintensive. Performing a migration with an inappropriate parameter may cause extensively long migration time or downtime. Defining this parameter is nontrivial since application workloads are generally unpredictable. This difficulty also makes it hard for cloud management systems to operate VM live migration automatically. In this paper, we propose the Memory-bound Pre-copy Live Migration (MPLM) mechanism that performs VM live migration without requiring the maximum tolerable downtime parameter. MPLM presents a new perspective of VM live migration, where the migration is performed based on the present state of VM computation without enforcing downtime constraints. During live migration, MPLM separates memory pages into dirty and non-dirty sets and transmits them in a multiplexing manner. MPLM has been implemented in a modified version of the QEMU-KVM software. Experiments have been conducted to evaluate MPLM performances against those of the traditional pre-copy mechanism. We performed a number of live migrations of VMs running four OpenMP NAS Parallel Benchmark programs. Experimental results show that MPLM is more efficient and easier to use than the pre-copy mechanism.