Προβολή Εργασίας
ΕΚΤΙΜΗΣΗ ΤΗΣ ΑΠΟΔΟΣΗΣ ΤΟΥ ΚΩΔΙΚΑ SPMV ΣΕ ΣΥΓΧΡΟΝΕΣ ΑΡΧΙΤΕΚΤΟΝΙΚΕΣ
Β. Καρακάσης, Γ. Γκούμας, Ν. Αναστόπουλος, Κ. Κούρτης, Ν. Κοζύρης
Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Η/Υ, Εθνικό Μετσόβιο Πολυτεχνείο |
ΣΥΝΤΟΜΗ ΠΕΡΙΛΗΨΗ
The performance attained by real scientific applications is usually much
lower than the peak performance that modern architectures can deliver, leaving
thereby unused a large amount of computing power. As architectures in the seek
of pure performance become more elaborate and complicated, simply
understanding the tradeoffs implied by the application itself is not adequate
to extract the maximum performance of the underlying machine. Thus, deeper
understanding of the architecture\'s performance tradeoffs is essential in order
to exploit as much of the delivered power as possible. Interconnection networks
play also an important role in the peak performance of large-scale computer
systems and should be considered, as well.
In this paper we are examining a well-known scientific computational kernel,
the Sparse Matrix-Vector Multiplication Kernel (spmv), which is used in a
variety of computational problems in physics, mechanical and chemical
engineering, etc. The goal of this paper is to unveil some subtle performance
issues of the spmv kernel, which arise in modern multiprocessor and multicore
architectures. Although this work focuses mainly on per-node optimization
issues, interconnection networks are also examined as they may drive specific
single-node optimizations, when clusters of modern multicore processors and
multiprocessors are considered.
Λέξεις Κλειδιά
Multiprocessor architectures, Multicore Architectures, Sparse Matrix-Vector Multiplication
|
|
