FT-OpenVZ: A Virtualized Approach to Fault-Tolerance in Distributed Systems

摘要

We present FT-OpenVZ, a full checkpointing and fault-tolerant solution for virtual private server (VPS) distributed computing. FT-OpenVZ extends Open VZ's VPS checkpointing to include full VPS checkpointing for MPI applications, including incremental file system checkpointing and user-assisted restart of checkpoints. With our solution, we extend the state of virtual machine/VPS fault-tolerance to any MPI-based distributed solution. By checkpointing all child processes, threads, and files within a distributed system, we provide a framework for future fault-tolerance work. For added resiliency to node failure, we include checkpoint replication and show that its use dramatically decreases the burden of checkpointing on network storage/centralized storage solutions. Using replication, FT-OpenVZ eliminates any need for network storage or centralized servers, reducing the impact of checkpointing on non-participating cluster nodes/users. Further, we show that by using replication our solution is scalable, where network storage and centralized server-based solutions are not. Our analysis is based on the NAS Parallel Benchmarks with cluster sizes up to 64 nodes. Using these benchmarks we examine the overhead of checkpointing with replication, demonstrating low overhead for virtualized checkpointing.