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:20191206T142452Z
LOCATION:Panorama 1
DTSTART;TZID=Europe/Stockholm:20180625T161000
DTEND;TZID=Europe/Stockholm:20180625T165000
UID:isc_hpc_ISC High Performance 2018_sess201_pap123@linklings.com
SUMMARY:Hans Meuer Award Finalist 1: Chebyshev Filter Diagonalization on M
 odern Manycore Processors and GPGPUs
DESCRIPTION:Hans Meuer Award Finalists, Research Paper\n\nHans Meuer Award
  Finalist 1: Chebyshev Filter Diagonalization on Modern Manycore Processor
 s and GPGPUs\n\nKreutzer, Ernst, Bishop, Fehske, Hager...\n\nChebyshev fil
 ter diagonalization is well established in quantum\nchemistry and quantum 
 physics to compute bulks of eigenvalues of\nlarge sparse matrices. Choosin
 g a block vector implementation, we\ninvestigate optimization opportunitie
 s on the new class of\nhigh-performance compute devices featuring both hig
 h-bandwidth and\nlow-bandwidth memory. We focus on the transparent access 
 to the full\naddress space supported by both architectures under considera
 tion:\nIntel Xeon Phi "Knights Landing" and Nvidia "Pascal"/"Volta."\nAfte
 r a thorough performance analysis of the single-device\nimplementations us
 ing the roofline model we propose two optimizations: (1)\nSubspace blockin
 g is applied for improved performance and data\naccess efficiency.  We als
 o show that it allows transparently\nhandling problems much larger than th
 e high-bandwidth memory without\nsignificant performance penalties. (2) Pi
 pelining of communication\nand computation phases of successive subspaces 
 is implemented to\nhide communication costs without extra memory traffic. 
  As an\napplication scenario we perform filter diagonalization studies for
 \ntopological quantum matter.  Performance numbers on up to 2048 nodes\nof
  the Oakforest-PACS and Piz Daint supercomputers are presented, achieving\
 nbeyond 500 Tflop/s for computing 100 inner eigenvalues of sparse\nmatrice
 s of dimension 4 billion.\n\nPasses: Conference Pass, Parallel Algorithms,
  Parallel Applications, Performance Analysis and Optimization\n\nTag: Conf
 erence Pass, Parallel Algorithms, Parallel Applications, Performance Analy
 sis and Optimization
URL:https://2018.isc-program.com/presentation/?id=pap123&sess=sess201
END:VEVENT
END:VCALENDAR