A Pipelined Multi-level Checkpoint Storage System for Virtual Cluster Checkpointing

Checkpoint-Restart (CR) is an important mechanism to provide fault tolerance for long-running applications in the cloud. Virtual Cluster (VC) checkpointing is the mechanism to perform CR operations on a cluster of Virtual Machines (VMs). However, saving of state of every VM in a VC directly to a shared storage, such as a SAN storage, simultaneously can cause high checkpoint overheads due to I/O contentions. This paper presents a novel Pipedlined Multi-Level Checkpoint Storage (PMLCS) system that allows users to schedule the saving of VM checkpoints on multiple storage devices to avoid I/O contention. We propose a novel conceptual diagram called the Backup Chain (BC) diagram to describe the scheduled saving and backing up of VM checkpoint files on various storage devices. We have implemented the PMLCS prototype and integrated it with two hypervisor-level VC checkpointing systems. We conducted experiments by checkpointing a VC running an MPI program from NAS parallel benchmark. Experimental results show that the PMLCS system using the BC diagram that represents a combination of SSD and SAN storage is an efficient and scalable checkpoint storage solution.