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:Analog 1\, 2
DTSTART;TZID=Europe/Stockholm:20180625T134900
DTEND;TZID=Europe/Stockholm:20180625T135300
UID:isc_hpc_ISC High Performance 2018_sess103_phd108@linklings.com
SUMMARY:(PhD13) Parallel Low Memory Footprint Eikonal Solver in Cardiovasc
 ular Applications
DESCRIPTION:PhD Forum\n\n(PhD13) Parallel Low Memory Footprint Eikonal Sol
 ver in Cardiovascular Applications\n\nGanellari, Haase\n\nThe basic equati
 ons in cardiac electrophysiology are the bidomain equations describing the
  intercellular and the extracellular electrical potential via a system of 
 two PDEs coupled non-linearly by a bunch of ODEs. Its difference, the tran
 s-membrane potential, is responsible for the excitation of the heart and i
 ts steepest gradients form an excitation wavefront propagating in time. Th
 is arrival time of the wavefront can be approximated by the Eikonal equati
 on with given heterogeneous and anisotropic velocity information.\n\nThe n
 umerical solution of the Eikonal equation follows the fast iterative metho
 d with its application for tetrahedral meshes. The accuracy of these simul
 ations is limited by unavailable patient specific conductivity data. \n\nT
 he human heart consists of various tissues with different conductivity par
 ameters.\nWe group these tissues into m different classes with its individ
 ual scaling parameter gamma_k yielding to the modified Eikonal equation wi
 th material domains where the velocity information  in each material domai
 n omega_k  is scaled by the parameter gamma_k. Now the activation time dep
 ends also on the scaling parameters gamma. \n\nOne chance to determine the
  scaling parameters suitably consists in comparing the Eikonal computed ac
 tivation sequence on the heart surface with the measured ECG on the torso 
 mapped onto this surface. The performance of the solver is crucial for the
  optimization performance especially when using an automatic differentiati
 on approach. Calculating the gradient via an analytic adjoint-state method
  is ongoing work.\n\nCurrently we have very performant parallel Eikonal so
 lvers which include shared memory parallelization and GPU parallelization 
 using CUDA. A low memory footprint solver where the memory access patterns
  are optimized is implemented on both solvers. The efficient implementatio
 n requires a local Gray-code numbering of edges in the tetrahedron and a b
 unch \nof bit shifts to assign the appropriate data. Numerical experiments
  on CPUs and GPUs show that the reduced memory footprint approach is faste
 r by 40\% than the original implementation.\n\nFor large scale problems, t
 he task based parallel model will run into difficulties: There might be no
 t enough (shared) memory on a single host or on a GPU, the computing power
  of a single compute unit is not sufficient, or the parallel efficiency is
  not satisfactory. In all cases, a distributed memory model is needed. Hen
 ce a coarser decomposition of the algorithm is needed, namely a domain dec
 omposition approach.\n\nWe present two different strategies on load balanc
 ing in CUDA in order to achieve to run the domain decomposition approach i
 n one GPU. \nThe first approach maps simply one sub-domain to one thread b
 lock. Its scaling improves with an increased number of sub-domains by redu
 cing the overall host synchronization together with the preallocation of t
 he global memory. The second approach takes better advantage of the GPU sh
 ared memory since it shares the workload of one sub-domain between many th
 read blocks exploiting in this way the total shared memory space.\n\nThe d
 omain decomposition approach is the first step towards the inter-process c
 ommunication implementation where the limitation of the global memory will
  be overcome completely by using multiple accelerator cards and cluster co
 mputing.\n\nPasses: Conference Pass, Exhibition Pass, HPC Accelerators, Pa
 rallel Algorithms\n\nTag: Conference Pass, Exhibition Pass, HPC Accelerato
 rs, Parallel Algorithms
URL:https://2018.isc-program.com/presentation/?id=phd108&sess=sess103
END:VEVENT
END:VCALENDAR