Fixed a bug in the counstrained source term under MPI #33
Conversation
9f3dd75 to
ca06b3b
Compare
|
@WaltherM Just updated. |
|
I see. I was just testing it at the momemt... I'll rebase and let you know about the status... |
|
can't test this due to #34 |
|
@norihiro-w Updated according to your comments. |
| @@ -26,6 +26,9 @@ | |||
| /*--------------------- MPI Parallel -------------------*/ | |||
| #if defined(USE_MPI) || defined(USE_MPI_PARPROC) || defined(USE_MPI_REGSOIL) | |||
| #include <mpi.h> | |||
| #ifdef OGS_FEM_IPQC | |||
There was a problem hiding this comment.
sorry after #35, we don't need this #ifdef anymore. SplitMPI_Communicator.h is always used if USE_MPI
|
@wenqing you can keep the original commit and remove the last commit including my comments. |
20e4429 to
ca06b3b
Compare
|
@norihiro-w Recovered. |
|
I will merge this after #35 |
|
@norihiro-w Thanks. @WaltherM should give a test before it is merged. |
|
will test after #35 is merged and this is rebased. |
|
#35 was merged |
|
@wenqing could you please help me to rebase this - I tried it on master, but ogs crashes with [node002:13167] *** An error occurred in MPI_Allreduce |
|
@WaltherM Rebased. If it still crashes, please send me your input files. |
| #else | ||
| MPI_Comm comm = comm_DDC; | ||
| #endif | ||
| MPI_Allreduce(&st_conditionst, &int_buff, 1, MPI_INT, MPI_SUM, comm); |
There was a problem hiding this comment.
@WaltherM collective communication function can’t be called here because that not call partitions satisfy the condition of (line 8044)
if (st_node.size() > 0 && (long) st_node.size() > i)
The means some compute cores skip the if-condition, while the other move inside the if-condition to call MPI_Allreduce, which leads to a dead lock here.
Due to this reason, I have to close this PR and want to discuss the issue with you later.
|
@wenqing ok, lets discuss this on Monday. |
As titled.