@device(postscript)
@libraryfile(Mathematics10)
@libraryfile(Accents)
@style(fontfamily=timesroman,fontscale=11)
@pagefooting(immediate, left "@c<Technical Report 1996>",
center "@c<Computer Science>",
right "@c<Carnegie Mellon>")
@heading(A Trace-Driven Comparison of Algorithms for
Multi-Process Prefetching and Caching)
@heading(CMU-CS-96-174)
@center(@b(Andrew Tomkins, R. Hugo Patterson, Garth Gibson))
@center(September 1996)
@center(FTP: CMU-CS-96-174.ps)
@blankspace(1)
@begin(text)

Recently two groups of researchers have proposed systems that exploit
application knowledge to improve I/O performance.  Both systems use
application knowledge to prefetch data thereby masking I/O latency and to
improve file buffer cache performance thereby avoiding slow I/O accesses
altogether.  Unfortunately, published studies of these two systems are
incomparable.  This technical report is a follow-on to a paper to appear in
OSDI96 comparing the TIP2 system of Patterson, Gibson, et al, and the
LRU-SP system of Cao, Felten, Karlin and Li, co-written by the two groups.
The OSDI paper considers the case of a single process with full advance
knowledge of requests.  In this technical report we consider multiple
processes, each of which has either full advance knowledge (complete hints) or
no advance knowledge (no hints).

Our results can be summarized as follows: the cost-benefit analysis of TIP2
allows better performance when optimal buffer allocation does not correspond
to process consumption rates.


@blankspace(2line)
@begin(transparent,size=10)
@b(Keywords:@ )@c<File systems, parallel I/O, prefetching, informed
caching, hints, TIP, LRU-SP>
@end(transparent)
@blankspace(1line)
@end(text)
@flushright(@b[(25 pages)])