Computer Science Colloquia
Wednesday, May 1, 2013
Ashwin Raghav Mohan Ganesh
Advisor: Andrew Grimshaw
Attending Faculty: Jack Stankovic (Chair), Kamin Whitehouse, and Marty Humphrey
10:00 AM, Rice Hall, Rm. 242
PhD Qualifying Exam Presentation
A Scalable Application-level Distributed Filesystem Notifications Solution
ABSTRACT
Filesystem event notifications help applications track file events.
Although most operating systems are provisioned to notify applications
of events on local file objects, there exists no scalable solution to
track events on files that reside on distributed filesystems. In this
paper, we introduce Rnotify, a scalable application level solution to
this problem.
Our system has four key properties. Firstly, the Rnotify API provides
application semantics nearly equivalent to the existing Linux
notifications API (Inotify). Linux applications can potentially use
Rnotify with minimal code changes. Secondly, by slightly modifying the
semantics of Inotify, Rnotify will dispatch event notifications under a
latency of 10 milliseconds for 85% of useful filesystem changes without
being subject to event queue overflows. Thirdly, applications that choose
to use Rnotify as a High throughput notification delivery system can do
so at low latency costs by distributing their notification processing.
Lastly, a clear decomposition of functionalities between different
components of Rnotify ensures that the system can accommodate different
workloads imposed by traditional HPC or modern Web Applications by
scaling specific components. We support our claims by testing Rnotify
with a GlusterFS deployment on an Amazon Web Services High Performance
Computing Cluster.