Performance and resource optimization is an important
research problem in data intensive distributed comput-
ing. We present a new batched stream processing model
that captures query correlations to expose I/O and com-
putation redundancies for optimizations. The model is
inspired by our empirical study on a trace from a pro-
duction large-scale data processing cluster, which reveals
significant redundancies caused by strong temporal and
spatial correlations among queries.
We have developed Comet, a query processing
system that embraces the batched stream processing
model for optimizations. We have integrated Comet with
DryadLINQ. With its roots in query optimizations for
database systems, Comet enables a set of new heuristics
and opportunities tailored for distributed computing in
DryadLINQ. Optimizations in Comet are effective. The
evaluation of a micro-benchmark on a 40-machine clus-
ter shows a 42% reduction in total machine time and over
40% reduction in total I/O. Our simulation on a real trace
covering over 19 million machine hours shows an esti-
mated I/O saving of over 50%.
Attachment | Size |
---|---|
He2008CometBatchedStreamProcessinginDataIntensiveDistributed.pdf | 579.16 KB |