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:20190627T201357Z
LOCATION:Panorama 1
DTSTART;TZID=Europe/Stockholm:20190619T144500
DTEND;TZID=Europe/Stockholm:20190619T151500
UID:isc_hpc_ISC High Performance 2019_sess221_inv_sp152@linklings.com
SUMMARY:Analysing and Tuning the Performance of Graph Processing Algorithm
s: a Statistical Modeling Approach
DESCRIPTION:Focus Session\nConference Pass, Graph Algorithms, HPC Accelera
tors, Parallel Algorithms, Performance Analysis and Optimization\n\nAnalys
ing and Tuning the Performance of Graph Processing Algorithms: a Statistic
al Modeling Approach\n\nVarbanescu\n\nLarge-scale and complex graph proces
sing applications form a challenging domain for high-performance computing
. Despite graph processing algorithms being considered parallelism-unfrien
dly, the use of parallel architectures like multi-core CPUs and GPUs hav
e proven revolutionary for these applications. However, analysing and mode
ling the performance of these algorithms on parallel platforms remains a c
hallenge: the tight dependencies between platform, algorithm, and dataset
are proven difficult to analytically determine, model, and feed back into
the algorithm design.\n\nIn this work, we present a comprehensive framewor
k for graph processing performance analysis, and further demonstrate its u
se for performance modeling and tuning. Our solution is based on a statist
ical approach, and combines efficient model training with accurate predict
ions. We are further able to use these predictions to improve algorithm ex
ecution. Finally, we present the performance analysis and tuning of two c
ase-studies (BFS and PageRank), and demonstrate how to use performance mod
eling to obtain better implementations, which clearly outperform state-of-
the-art implementations.
URL:https://2019.isc-program.com/presentation/?id=inv_sp152&sess=sess221
END:VEVENT
END:VCALENDAR