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:20210916T132450Z
LOCATION:Jean-Jacques Rousseau
DTSTART;TZID=Europe/Stockholm:20210707T110000
DTEND;TZID=Europe/Stockholm:20210707T113000
UID:submissions.pasc-conference.org_PASC21_sess161_msa261@linklings.com
SUMMARY:Performance Improvements in Multilevel DDM through Novel Features 
 in Sparse Direct Solvers
DESCRIPTION:Minisymposium\n\nPerformance Improvements in Multilevel DDM th
 rough Novel Features in Sparse Direct Solvers\n\nButtari, Darrigrand, Joli
 vet\n\nIn the context of the EoCoE project, and, in particular, for the si
 mulation of wind farms, many efforts are devoted to the resolution of larg
 e structural mechanics problems which require the use of robust solvers. S
 parse direct methods represent a reliable choice for handling moderate siz
 e problems of this type but their complexity results in limited scalabilit
 y when moving to larger ones. As an alternative, multilevel domain decompo
 sition methods (DDM) can be used to build robust preconditioners for itera
 tive solvers. These methods rely on a spectral coarse space to damp the lo
 w-frequency errors and employ sparse direct solvers to handle local subpro
 blems. Although this approach provides a good trade-off between robustness
  and scalability the direct solver involved in the resolution on the subdo
 mains remains the bottleneck. To tackle this issue, we have investigated t
 he use of recent advances in sparse direct solvers, namely, advanced multi
 threading techniques and Block Low Rank (BLR) approximations which reduce 
 the asymptotic complexity. We have implemented this approach using the HPD
 DM and MUMPS packages and evaluated its effectiveness on a large heterogen
 eous linear elastic problem. Experimental results show that, although the 
 loss of accuracy resulting from BLR can be reliably controlled, this featu
 re must be used with caution because it may deeply affect the convergence 
 of the preconditioned iterative method. Additionally, because BLR performs
  best on relatively large-size subproblems, efficient multithreading is es
 sential to achieve high performance. In conclusion, we show that, through 
 fine-tuning of the parameters, the overall time to solution can be conside
 rably reduced.\n\nDomain: CS and Math, Emerging Applications, Solid Earth 
 Dynamics, Engineering
END:VEVENT
END:VCALENDAR