[cig-commits] r22682 - seismo/3D/SPECFEM3D_GLOBE/trunk/src/specfem3D

dkomati1 at geodynamics.org dkomati1 at geodynamics.org
Thu Jul 25 16:53:04 PDT 2013

Author: dkomati1
Date: 2013-07-25 16:53:04 -0700 (Thu, 25 Jul 2013)
New Revision: 22682

added a comment for Daniel about the fact that buffers to undo attenuation should remain on the host in the case of GPU_MODE, they would not fit on the device, and the host has more free memory anyway

Modified: seismo/3D/SPECFEM3D_GLOBE/trunk/src/specfem3D/specfem3D.F90
--- seismo/3D/SPECFEM3D_GLOBE/trunk/src/specfem3D/specfem3D.F90	2013-07-25 23:25:48 UTC (rev 22681)
+++ seismo/3D/SPECFEM3D_GLOBE/trunk/src/specfem3D/specfem3D.F90	2013-07-25 23:53:04 UTC (rev 22682)
@@ -2501,6 +2501,15 @@
 ! in exact undoing of attenuation
     undo_att_sim_type_3 = .true.
+!! DK DK to Daniel, July 2013: in the case of GPU_MODE it will be *crucial* to leave these arrays on the host
+!! i.e. on the CPU, in order to be able to use all the (unused) memory of the host for them, since they are
+!! (purposely) huge and designed to use almost all the memory available (by carefully optimizing the
+!! value of NT_DUMP_ATTENUATION); when writing to these buffers, it will then be OK to use non-blocking writes
+!! from the device to the host, since we do not reuse the content of these buffers until much later, in a second part
+!! of the run; however when reading back from these buffers, the reads from host to device should then be blocking
+!! because we then use the values read immediately (one value at a time, but to reduce the total number of reads
+!! across the PCI-Express bus we could / should consider reading them 10 by 10 for instance (?) if that fits
+!! in the memory of the GPU
     if( ier /= 0 ) call exit_MPI(myrank,'error allocating b_displ_crust_mantle_store_buffer')

More information about the CIG-COMMITS mailing list