[Einstein Toolkit] #2177: CarpetInterp tree_search fails when using Carpet with manual topology specified
#2177: CarpetInterp tree_search fails when using Carpet with manual topology specified ---------------------------------+-------------------- Reporter: yosef@… | Type: defect Status: new | Priority: unset Milestone: | Component: Other Version: development version | Keywords: ---------------------------------+-------------------- While running a unigrid simulation in Carpet, I ran into an issue with interpolation. The CarpetInterp tree_search algorithm fails when I used Carpet::processor_topology = "manual"
Is this expected? Setting CarpetInterp::tree_search = false seems to make the problem go away.
This is the reported error when tree_search = yes
Inconsistent search result from find_location_tree for interpolation point #1626 at [0.043412,-0.246202,3.0151e-17] of patch #0 is not on any component
I attached a parfile showing the issue.
#2177: CarpetInterp tree_search fails when using Carpet with manual topology specified ----------------------+--------------------------------- Reporter: yosef@… | Owner: (none) Type: defect | Status: new Priority: unset | Milestone: Component: Other | Version: development version Resolution: | Keywords: ----------------------+--------------------------------- Changes (by yosef@…):
* Attachment "simple.par" added.
parfile that triggers tree_search error
#2177: CarpetInterp tree_search fails when using Carpet with manual topology specified ----------------------+--------------------------------- Reporter: yosef@… | Owner: (none) Type: defect | Status: new Priority: unset | Milestone: Component: Other | Version: development version Resolution: | Keywords: ----------------------+---------------------------------
Comment (by Erik Schnetter):
This is not expected, but understandable. Apparently, the manual process decomposition does not properly initialize the data structure upon which the tree search relies. Given that a manual specification is rarely used (is it?), and that there is a work-around, this might not be corrected in the near term.
The tree search speeds up interpolation when there are many processes. Is interpolation a bottleneck in this simulation? How many processes are you using?
#2177: CarpetInterp tree_search fails when using Carpet with manual topology specified ----------------------+--------------------------------- Reporter: yosef@… | Owner: (none) Type: defect | Status: new Priority: unset | Milestone: Component: Other | Version: development version Resolution: | Keywords: ----------------------+---------------------------------
Comment (by yosefz@…):
I would expect that we wouldn't use more than 512 cores. What would be needed to properly initialize the appropriate data structures?
#2177: CarpetInterp tree_search fails when using Carpet with manual topology specified ----------------------+--------------------------------- Reporter: yosef@… | Owner: (none) Type: defect | Status: new Priority: unset | Milestone: Component: Other | Version: development version Resolution: | Keywords: ----------------------+---------------------------------
Comment (by Erik Schnetter):
I assume that the process regions are initialized, but the superregions are not initialized. There should be one superregion, defining a tree structure for the actual regions.
Is interpolation actually a bottleneck in the simulation?
#2177: CarpetInterp tree_search fails when using Carpet with manual topology specified ----------------------+--------------------------------- Reporter: yosef@… | Owner: (none) Type: defect | Status: new Priority: unset | Milestone: Component: Other | Version: development version Resolution: | Keywords: ----------------------+---------------------------------
Comment (by yosef@…):
Interpolation is not yet a bottleneck. The code is not yet used in any production setting. I found this issue while debugging an unrelated problem.
trac@lists.einsteintoolkit.org
-
Einstein Toolkit