Fault Tolerance in Large Clusters
The increased reliance on supercomputing clusters based on
commercial off-the-shelf (COTS) parts has created a growing need for
lightweight fault tolerance. As clusters grow from hundreds of
processors to thousands, and even tens of thousands of processors,
the mean time to failure (MTTF) drops from a few days to a less than
an hour for the entire cluster (assuming that the failure rate is
geometric and that the reliability of a single-processor system is
an optimistic 99.99% or MTTF = 10,000 hours = 1.14 years).
Due to the size of the clusters in question, resource constraints
make traditional fault-tolerance techniques based on active
replications impractical. However, rollback-recovery protocols
provide an attractive low-overhead solution for making such
clusters resilient to crash failures. These protocols save
information during failure-free execution that can be used during
recovery to restore a crashed process to its pre-crash state.
This joint project, led by Lorenzo Alvisi and Harrick Vin from the
University of Texas at Austin, builds on their prior work on Egida,
an object-oriented toolkit designed to support transparent rollback
recovery and low-overhead fault tolerance.
