abstract

    We have designed and implemented the Warp Transactional
Filesystem, a distributed filesystem for storing large
data.  It supports atomic file operations, and has
better aggregate throughput, latency, and fault
tolerance properties than existing publically available
distributed file systems.
    While previous distributed filesystems have many of the
same features as WTF, they do not provide strong enough
consistency guarantees to implement useful features such
as atomic file append.  Furthermore, existing designs
rely on a single master to store metadata in memory,
resulting in non-uniform hardware requirements, weak fault
tolerance properties, filesystem size limitations, 
and performance bottlenecks.
    In this paper we present the transactional API, 
design and implementation details, and performance
evaluation of the Warp Transactional Filesystem.