Decomposition stabilization#27
Conversation
… preferring decompositions with singleton dimensions and an option to choose between different equally-weighted decompositions.
…est using manual overrides - check verbose output for options
…the stabilization patch has been driven into production
|
To be merged with fmihpc/vlasiator#821 |
markusbattarbee
left a comment
markusbattarbee
left a comment
There was a problem hiding this comment.
See Vlasiator PR: recommend consolidating values as uint64_t.
Also, I think we should have one legacy decomposition calculation function which is just the old type turned into deterministic with uint64_t's, and then a new one we can use in the future which is used with stored decomposition in the file.
ykempf
left a comment
ykempf
left a comment
There was a problem hiding this comment.
I owuld advocate to make dd domaindecomposition or so, and test-dd vs ddtest more consistent. But well...
| int myRank, MPI_flag; | ||
| MPI_Initialized(&MPI_flag); | ||
| if(MPI_flag){ | ||
| MPI_Comm_rank(MPI_COMM_WORLD, &myRank); // TODO allow for separate communicator |
There was a problem hiding this comment.
That will be needed in conjunction with Jaro's branch, right?
| subarrayStart[i] = 0; | ||
|
|
||
| std::array<int,3> swappedStorageSize = storageSize; | ||
| std::array<int,3> swappedStorageSize = {(int)storageSize[0],(int)storageSize[1],(int)storageSize[2]}; |
There was a problem hiding this comment.
So this makes an actual difference beyond making the type conversion extra clear??
There was a problem hiding this comment.
Gets rid of compile warnings (storageSize is of type FsSize_t)
| // 1) Task grid in the MPI_Cartcomm | ||
| std::array<int, 3> ntasks; //!< Number of tasks in each direction | ||
| std::array<int, 3> taskPosition; //!< This task's position in the 3d task grid | ||
| std::array<Task_t, 3> ntasks; //!< Number of tasks in each direction |
There was a problem hiding this comment.
This is historic but could be named slightly more explicitly?
…asksPerDim instead for actual clarity.
3 participants
Making the fsgrid decomposition deterministic, with manual support (e.g. for using existing decompositions)