Reliable MPI-IO through Layout-Aware Replication

The current deployment of petascale systems and the promise of future exascale systems have created unprecedented challenges in how to manage failures in such systems. While many parallel file systems provide some sort of redundancy mechanism to cope with failures, such systems rely heavily on a hardware-based solution such as RAID. In this paper, we propose a block replication approach to store data redundantly. The approach does not depend on file system fault-tolerance mechanisms. Rather, the approach replicates each file block transparently within MPI-IO, using replicationaware datatypes. File striping information is used to place blocks from each replica in a separate storage node. We have implemented this replication mechanism in the MPI-IO layer. Our experimental results using a microbenchmark and real MPI-IO applications with PVFS and Lustre demonstrate that block replication in MPI-IO can be achieved transparently.