Michael

This is likely caused by your grid structure. Some refinement levels
probably have several components (maybe one component surrounding each
black hole in a binary system), and the code in question probably has not
yet been extended to handle this case.

The loop is there to set up certain internal variables correctly. This
particular routine would work fine with multiple local components, but
others will not. You see e.g. that the signature of the
function Device_CopyPreSync accepts a variable index, refinement level, and
time level as input. All other indices must therefore be trivial (multigrid
level, component, map index) since the called routine has no way of
determining them. It should be relatively straightforward but tedious to
extend the API to handle these as well.

-erik


On Thu, Dec 17, 2015 at 2:04 PM, Michael Clark <[email protected]>
wrote:

> Hello,
>
> I'm tripping the assert(nlcs==1) in Accelerator_PreSync when I try to run
> with a single MPI process.
>
> In particular, I am working on a system with 2 MPI sockets per node (16
> cores per node) and would like to try using a single MPI socket (either
> with all 16 cores, or with only 8 cores).  This is desired because I'm
> trying to track down a discrepancy in data produced by GPU code and CPU
> code--we think that data in the regions that must be communicated by MPI
> might not be correct for GPU code, so we would like to test this
> possibility by running with only one MPI process.
>
> We gather that nlcs means the number of local components, but we're
> puzzled that there is a loop over components immediately after this
> assertion.
>
> I see this with various resolutions, with total required memory (as
> reported by carpet) ranging from 1 to 4 GB.
>
> Is this a grid structure issue, a domain size issue, or something else?
> Is this assert necessary and required given the loop that occurs
> immediately after it?  What else can I do to resolve this problem and
> perform the test I would like to perform?
>
> Thanks,
> Michael Clark
>
> _______________________________________________
> Users mailing list
> [email protected]
> http://cactuscode.org/mailman/listinfo/users
>
>


-- 
Erik Schnetter <[email protected]>
http://www.perimeterinstitute.ca/personal/eschnetter/
_______________________________________________
Users mailing list
[email protected]
http://cactuscode.org/mailman/listinfo/users

Reply via email to