Speaker
Mr
Fabrizio Furano
(INFN sez. di Padova)
Description
The latencies induced by network communication often play a big role in reducing the
performances of systems which access big amounts of data in a distributed
environment. The problem is present in Local Area Networks, but in Wide Area Networks
is much more evident. It is generally perceived as a critical problem which makes
very difficult to get access to remote data. However, a more detailed analysis on the
access pattern of the involved applications can be used to understand the
characteristics of the stream of the data requests, and develop techniques to
optimize it. This work started from the analysis of the access patterns of the BaBar
experiment's physics analysis data, but the methods and the results can be applied in
other computing environments as well. We show how the exploit of caching and
asynchronous prefetching techniques is able to enhance the performances of such kind
of applications in Local Area Networks, and is able to lower the total latencies for
Wide Area Networks data access of an order of magnitude. Moreover, the ability to
process file open requests in parallel can be a very interesting performance
enhancement for applications which need access to many files at once. Such general
techniques have been implemented in the client side of the xrootd data access system,
which, for its performances and its fault tolerant architecture, showed itself as an
ideal testbed for such kind of enhancements.
Primary author
Mr
Fabrizio Furano
(INFN sez. di Padova)
