BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20220812T074335Z
LOCATION:Singapore Room
DTSTART;TZID=Europe/Stockholm:20220629T110000
DTEND;TZID=Europe/Stockholm:20220629T113000
UID:submissions.pasc-conference.org_PASC22_sess156_msa117@linklings.com
SUMMARY:RACE: Speeding up Spectral Clustering and Iterative Solvers using
Level-Based Parallelization and Blocking Techniques
DESCRIPTION:Minisymposium\n\nRACE: Speeding up Spectral Clustering and Ite
rative Solvers using Level-Based Parallelization and Blocking Techniques\n
\nAlappat, Pasadakis, Hager, Schenk, Wellein\n\nClustering and linear iter
ative solvers are indispensable for large-scale simulations in the area of
machine learning and mathematical optimization. In this talk, we present
methods to accelerate existing spectral clustering techniques and solvers
by using the concept of levels as developed in the context of our RACE (Re
cursive Algebraic Coloring Engine) library framework. Levels are construct
ed using breadth-first-search (BFS) on the graph related to the underlying
sparse matrix. These levels are then used to implement cache blocking of
the matrix elements for high spatial and temporal reuse. The approach find
s its use in kernels like sparse-matrix-power vector multiplication and Ch
ebyshev polynomial iterations, which perform repetitive back-to-back spars
e-matrix-vector multiplication (SpMV)-type iterations without global synch
ronizations in between. The method is highly effective and achieves perfor
mance levels of 50-100 GF/s on a single modern Intel or AMD multicore chip
, providing speedups of typically 2x - 4x compared to a highly optimized c
lassical SpMV implementation.
After introducing the optimizati
on strategy, we discuss the application of the method to a highly scalable
power iteration graph clustering (PIC) algorithm and the coupling of RACE
with the Trilinos framework to speed up a class of linear iterative solve
rs commonly found in optimization algorithms.\n\nDomain: Computer Science
and Applied Mathematics
END:VEVENT
END:VCALENDAR