Odd MPI IO bug in Open–MPI
Quite often working with research code leads you to find the unusual edge cases where even well tested libraries break down a bit. This example that we ran across with the Open-MPI parallel library is pretty typical. We wanted to use MPI-IO to write an array that was distributed across multiple processors, with each processor holding a fraction of the array. Because of how we were using the arrays, the array on each processor had a strip of "guard" cells along each edge that were used to exchange information with the neighbouring processor and these had to be clipped off before the array was written. MPI makes this very easy to achieve using MPI types. (This example is given in Fortran, because that was the language that we encountered the problem in. C would have the same problems)
First you create a type representing the total array using MPI_Type_create_subarray
sizes = (/nx_global, ny_global/) subsizes = (/nx_local, ny_local/) starts = (/x_cell_min, y_cell_min/) CALL MPI_TYPE_CREATE_SUBARRAY(ndims, sizes, subsizes, starts, & MPI_ORDER_FORTRAN, MPI_DOUBLE_PRECISION, global_type, ierr)
This specifies an array that's globally 1:nx_global x 1:ny_global, and locally 1:nx_local x 1:ny_local. The starts array specifies where the local part of the global array starts in the global array, and depends on how you split your global array over processors. Then you pass this type as a fileview to MPI_File_set_viewto tell MPI that this is how data is arranged across your processors.
The actual data is in an array one bigger on each end (0:nx_local+1x 0:ny_local+1), so we need another type representing how to cut off those additional cells. That's MPI_Type_create_subarray again
sizes = (/nx_local+2, ny_local+2/) subsizes = (/nx_local, ny_local/) starts = (/1, 1/) CALL MPI_TYPE_CREATE_SUBARRAY(ndims, sizes, subsizes, starts, & MPI_ORDER_FORTRAN, MPI_DOUBLE_PRECISION, core_type, ierr)
When you pass this as the datatype to a call to MPI_File_write or MPI_File_write_allyou pass MPI only the 1:nx_localx 1:ny_local array that you promised it when you called MPI_File_set_view. The final result will be an array 1:nx_globalx 1:ny_globalno matter how many processors you run the code on.
The problem was that it wasn't working. When we ran the code we found that everything worked as expected on files < 512MB/processor in size, but when we got beyond that the files were always systematically smaller than expected. They weren't a fixed size, but they were always smaller than expected. As we always advise other people to do we started from the assumption that we had made a mistake somewhere, so we went over our IO code and our MPI types. They all appeared normal, so we started taking parts out of our code. After removing a few bits, we found that the critical element was using the MPI derived type to clip out the guard cells from the local arrays. If we just wrote an entire array using primitive MPI types the problem didn't occur. This was about the point where it started to look like it might, just possibly, be an MPI error.
Next, we created the simplest possible test case in Fortran that replicated the problem and it turned out to be very simple indeed. Just create the types for the filetype to MPI_File_set_view and the datatype to MPI_File_write and write an array larger than 512MB/processor. It even failed if you just coded it up for a single processor! It was unlikely at this stage that we'd just made a mistake in our trivial example code. As a final check, we then replicated it in C and found the same problem happened there. Finally, with working examples and some evidence that the problem wasn't in our code, we contacted the Open-MPI mailing list. Within a few hours, it was confirmed that we'd managed to find an edge case in the Open-MPI library and they'd created patches for the new versions of Open-MPI.
There are a couple of take away messages from this
- If you are using MPI-IO and are finding problems when files get larger than 512MB/processor you might need to update your MPI installation
- Sometimes there really are bugs in well tested and widely used libraries
- It's a good idea to be as sure as possible that you aren't making a mistake before assuming that you've found one.