nek5000 / parrsb Goto Github PK

Tested on BGQ using fparRSB_partMesh()

The current implentation do not guarantee that the partitions are contiguous.

Currently, the following functions are not valgrind clean:

• rcb_local
• findSegments
• mg_setup

• Remove gmsh reader from src
• Add gmsh reader to example
• Output results as vtk

Run gencon + parRSB to generate a ma2 file.

POSIX has reserved these names.

We can get rid of this O(n^2) loop:
https://github.com/Nek5000/parRSB/blob/master/src/con-periodic.c#L188

Currently we're using a fixed iteration counts which is not robust and efficient.

Describe the bug
Sometime in the last month, something broke in parSB related to how the connectivity works for serial cases. When I run a case in serial, I get a negative Jacobian error for every element in the mesh, but this error disappears and the case runs as normal if I run in parallel.

The reason I'm sure the error lies somewhere in parSB is because if you pre-generate the .co2 file, there is no error. So something is wrong with the automatic .co2 file generation for serial problems.

To Reproduce
I've attached a few input files that run with nekRS for you to replicate the problem. Download these files and run the following:

This fails with negative Jacobian errors:

nrsmpi pyramid 1 

This works:

nrsmpi pyramid 2

This also works:

gencon
nrsmpi pyramid 1

Expected behavior
Cases should be able to run in serial.

• Remove type from struct rcb_element and struct rsb_element
• Remove globalId0 from struct rsb_element
• Replace GenmapMalloc by tmalloc when allocating structs
• Replace GenmapMalloc by buffer_reserve when allocating arrays
• Replace GenmapFree by free and buffer_free
• Replace GenmapX by gslib types where X is int, long, uint, ulong.

Currently, we have hard coded gs_crystal in genmap-laplacian.c and gs_pairwise in
genmap-quality.c. We should select these method automatically. May be we do this once
at the top level ?

Our current bucket sort of the Fielder vector results in unbalanced bin sizes. Hence, we may run out of memory on some ranks.