Quantcast

Hyperslab collective writing crash with HDF5 1.8.16

classic Classic list List threaded Threaded
6 messages Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Hyperslab collective writing crash with HDF5 1.8.16

victorsv
Hi all,

I'm developing XH5For, an OO fortran light-weight XDMF/HDF5 layer, and now I would like to test its scalability, but I'm lost with an issue that i'm getting.

I'm going to try to explain it as best as I can.

I'm performing some weak scalability tests on Marenostrum III (GPFS file system) using Collective Writing with the Contiguous HyperSlab strategy. I'm running tests with 1, 16, 32, 64, 128, 256, 512, 1024 and 2048 MPI tasks.

All seems to work as expected except for the 2048 MPI tasks test, where I think I'm getting an MPI interlock (it's still running without doing anything till the job time exceeds and the job is killed).

After that, i try to reproduce an error with a number of MPI tasks between 1024-2048, and I could get the following error message while launching a smaller size job, with 1164 MPI tasks:

HDF5-DIAG: Error detected in HDF5 (1.8.16) MPI-process 1009:
  #000: H5Dio.c line 271 in H5Dwrite(): can't prepare for writing data
    major: Dataset
    minor: Write failed
  #001: H5Dio.c line 352 in H5D__pre_write(): can't write data
    major: Dataset
    minor: Write failed
  #002: H5Dio.c line 789 in H5D__write(): can't write data
    major: Dataset
    minor: Write failed
  #003: H5Dmpio.c line 529 in H5D__contig_collective_write(): couldn't finish shared collective MPI-IO
    major: Low-level I/O
    minor: Write failed

I have used the following libraries/versions during compilation stage:
Here you can see how I open the HDF5 file for Collective Writing:
https://github.com/victorsndvg/XH5For/blob/master/src/lib/hdf5_handler/hdf5_handler.f90#L531

Note: ENABLE_MPI, ENABLE_HDF5. ENABLE_PARALLEL_HDF5 definition flags are enabled

Could anyone give me some ligth about this?

I would greatly appreciate your help!

Thank you in advance,
Víctor.



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Hyperslab collective writing crash with HDF5 1.8.16

victorsv
Hi again,

In order to help to understand what is happening,  I've implemented and example using the HDF5 library.

You can see it in the following link:
https://github.com/victorsndvg/XH5For/blob/master/src/examples/hdf5_performance_test/ch_unstructured_hexahedron_perf.f90

I've compiled and launched it reproducing the previous mail and I'm getting the same behavior/errors with the new code.

I also compiled and linked against HDF5 1.8.17 by myself with the same results.

I'm not using any tunning hint ... this could be the problem?

Thanks!
Víctor



2016-07-19 11:08 GMT+02:00 victor sv <[hidden email]>:
Hi all,

I'm developing XH5For, an OO fortran light-weight XDMF/HDF5 layer, and now I would like to test its scalability, but I'm lost with an issue that i'm getting.

I'm going to try to explain it as best as I can.

I'm performing some weak scalability tests on Marenostrum III (GPFS file system) using Collective Writing with the Contiguous HyperSlab strategy. I'm running tests with 1, 16, 32, 64, 128, 256, 512, 1024 and 2048 MPI tasks.

All seems to work as expected except for the 2048 MPI tasks test, where I think I'm getting an MPI interlock (it's still running without doing anything till the job time exceeds and the job is killed).

After that, i try to reproduce an error with a number of MPI tasks between 1024-2048, and I could get the following error message while launching a smaller size job, with 1164 MPI tasks:

HDF5-DIAG: Error detected in HDF5 (1.8.16) MPI-process 1009:
  #000: H5Dio.c line 271 in H5Dwrite(): can't prepare for writing data
    major: Dataset
    minor: Write failed
  #001: H5Dio.c line 352 in H5D__pre_write(): can't write data
    major: Dataset
    minor: Write failed
  #002: H5Dio.c line 789 in H5D__write(): can't write data
    major: Dataset
    minor: Write failed
  #003: H5Dmpio.c line 529 in H5D__contig_collective_write(): couldn't finish shared collective MPI-IO
    major: Low-level I/O
    minor: Write failed

I have used the following libraries/versions during compilation stage:
Here you can see how I open the HDF5 file for Collective Writing:
https://github.com/victorsndvg/XH5For/blob/master/src/lib/hdf5_handler/hdf5_handler.f90#L531

Note: ENABLE_MPI, ENABLE_HDF5. ENABLE_PARALLEL_HDF5 definition flags are enabled

Could anyone give me some ligth about this?

I would greatly appreciate your help!

Thank you in advance,
Víctor.




_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Hyperslab collective writing crash with HDF5 1.8.16

Scot Breitenfeld
I suspect that it is an issue with the impi implementation. I have not been able to compile your code on our systems due to various compilation errors (this is using a different compiler as I don’t have access to impi). Have you tried using mpich (or some other mpi implementation) instead to see if that works? Also, have you tried using 1.10.0?

Scot


On Jul 23, 2016, at 1:12 PM, victor sv <[hidden email]> wrote:

Hi again,

In order to help to understand what is happening,  I've implemented and example using the HDF5 library.

You can see it in the following link:
https://github.com/victorsndvg/XH5For/blob/master/src/examples/hdf5_performance_test/ch_unstructured_hexahedron_perf.f90

I've compiled and launched it reproducing the previous mail and I'm getting the same behavior/errors with the new code.

I also compiled and linked against HDF5 1.8.17 by myself with the same results.

I'm not using any tunning hint ... this could be the problem?

Thanks!
Víctor



2016-07-19 11:08 GMT+02:00 victor sv <[hidden email]>:
Hi all,

I'm developing XH5For, an OO fortran light-weight XDMF/HDF5 layer, and now I would like to test its scalability, but I'm lost with an issue that i'm getting.

I'm going to try to explain it as best as I can.

I'm performing some weak scalability tests on Marenostrum III (GPFS file system) using Collective Writing with the Contiguous HyperSlab strategy. I'm running tests with 1, 16, 32, 64, 128, 256, 512, 1024 and 2048 MPI tasks.

All seems to work as expected except for the 2048 MPI tasks test, where I think I'm getting an MPI interlock (it's still running without doing anything till the job time exceeds and the job is killed).

After that, i try to reproduce an error with a number of MPI tasks between 1024-2048, and I could get the following error message while launching a smaller size job, with 1164 MPI tasks:

HDF5-DIAG: Error detected in HDF5 (1.8.16) MPI-process 1009:
  #000: H5Dio.c line 271 in H5Dwrite(): can't prepare for writing data
    major: Dataset
    minor: Write failed
  #001: H5Dio.c line 352 in H5D__pre_write(): can't write data
    major: Dataset
    minor: Write failed
  #002: H5Dio.c line 789 in H5D__write(): can't write data
    major: Dataset
    minor: Write failed
  #003: H5Dmpio.c line 529 in H5D__contig_collective_write(): couldn't finish shared collective MPI-IO
    major: Low-level I/O
    minor: Write failed

I have used the following libraries/versions during compilation stage:
Here you can see how I open the HDF5 file for Collective Writing:
https://github.com/victorsndvg/XH5For/blob/master/src/lib/hdf5_handler/hdf5_handler.f90#L531

Note: ENABLE_MPI, ENABLE_HDF5. ENABLE_PARALLEL_HDF5 definition flags are enabled

Could anyone give me some ligth about this?

I would greatly appreciate your help!

Thank you in advance,
Víctor.



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Hyperslab collective writing crash with HDF5 1.8.16

victorsv
Hi Scott,

thanks for your response.

I've tried with GNU compilers, OpenMPI and HDF5 (1.8.15, 1.8.16, 1.18.17), but in a local scope or small cluster, not in Marenostrum.

Ok, I'm going to try to compile all this tools to reproduce this environment and I will tell you if there are some news.

Can you report the compilation errors you are talking about? I can help you in this point or fix any issue!

Thank you again for your effort!
Víctor.

2016-07-26 16:41 GMT+02:00 Scot Breitenfeld <[hidden email]>:
I suspect that it is an issue with the impi implementation. I have not been able to compile your code on our systems due to various compilation errors (this is using a different compiler as I don’t have access to impi). Have you tried using mpich (or some other mpi implementation) instead to see if that works? Also, have you tried using 1.10.0?

Scot


On Jul 23, 2016, at 1:12 PM, victor sv <[hidden email]> wrote:

Hi again,

In order to help to understand what is happening,  I've implemented and example using the HDF5 library.

You can see it in the following link:
https://github.com/victorsndvg/XH5For/blob/master/src/examples/hdf5_performance_test/ch_unstructured_hexahedron_perf.f90

I've compiled and launched it reproducing the previous mail and I'm getting the same behavior/errors with the new code.

I also compiled and linked against HDF5 1.8.17 by myself with the same results.

I'm not using any tunning hint ... this could be the problem?

Thanks!
Víctor



2016-07-19 11:08 GMT+02:00 victor sv <[hidden email]>:
Hi all,

I'm developing XH5For, an OO fortran light-weight XDMF/HDF5 layer, and now I would like to test its scalability, but I'm lost with an issue that i'm getting.

I'm going to try to explain it as best as I can.

I'm performing some weak scalability tests on Marenostrum III (GPFS file system) using Collective Writing with the Contiguous HyperSlab strategy. I'm running tests with 1, 16, 32, 64, 128, 256, 512, 1024 and 2048 MPI tasks.

All seems to work as expected except for the 2048 MPI tasks test, where I think I'm getting an MPI interlock (it's still running without doing anything till the job time exceeds and the job is killed).

After that, i try to reproduce an error with a number of MPI tasks between 1024-2048, and I could get the following error message while launching a smaller size job, with 1164 MPI tasks:

HDF5-DIAG: Error detected in HDF5 (1.8.16) MPI-process 1009:
  #000: H5Dio.c line 271 in H5Dwrite(): can't prepare for writing data
    major: Dataset
    minor: Write failed
  #001: H5Dio.c line 352 in H5D__pre_write(): can't write data
    major: Dataset
    minor: Write failed
  #002: H5Dio.c line 789 in H5D__write(): can't write data
    major: Dataset
    minor: Write failed
  #003: H5Dmpio.c line 529 in H5D__contig_collective_write(): couldn't finish shared collective MPI-IO
    major: Low-level I/O
    minor: Write failed

I have used the following libraries/versions during compilation stage:
Here you can see how I open the HDF5 file for Collective Writing:
https://github.com/victorsndvg/XH5For/blob/master/src/lib/hdf5_handler/hdf5_handler.f90#L531

Note: ENABLE_MPI, ENABLE_HDF5. ENABLE_PARALLEL_HDF5 definition flags are enabled

Could anyone give me some ligth about this?

I would greatly appreciate your help!

Thank you in advance,
Víctor.



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Hyperslab collective writing crash with HDF5 1.8.16

Scot Breitenfeld
Using hdf5 1.8.17, I ran your program on a lustre system using Cray’s mpi wrappers for the intel compiler on 3072, 1536, 1200, 2016 processes with no issues (it completed in under 20s). If you have not already done so, you might want to make sure  the hdf5 parallel test programs all passed.

FYI, I managed to get it to compile by adding the static flag for the compiler, I had to hard code the option in cmake. Here is what I used to compile it,

cmake \
    -D CMAKE_C_FLAGS:STRING="-g" \
    -D CMAKE_Fortran_FLAGS:STRING="-g -static" \
    -D CMAKE_BUILD_TYPE:STRING="Debug" \
    -D XH5For_ENABLE_HDF5=ON \
    -D XH5For_ENABLE_MPI=ON \
    -D XH5For_ENABLE_TESTS=ON \
    -D XH5For_ENABLE_EXAMPLES=ON \
    -D BUILD_SHARED_LIBS=OFF \
    -D HDF5_INCLUDE_DIRS:PATH=$HDF5_DIR/include \
    -D HDF5_LIBRARY_DIRS:PATH=$HDF5_DIR/lib \
    $XH5For
 
Scot


On Jul 26, 2016, at 11:09 AM, victor sv <[hidden email]> wrote:

Hi Scott,

thanks for your response.

I've tried with GNU compilers, OpenMPI and HDF5 (1.8.15, 1.8.16, 1.18.17), but in a local scope or small cluster, not in Marenostrum.

Ok, I'm going to try to compile all this tools to reproduce this environment and I will tell you if there are some news.

Can you report the compilation errors you are talking about? I can help you in this point or fix any issue!

Thank you again for your effort!
Víctor.

2016-07-26 16:41 GMT+02:00 Scot Breitenfeld <[hidden email]>:
I suspect that it is an issue with the impi implementation. I have not been able to compile your code on our systems due to various compilation errors (this is using a different compiler as I don’t have access to impi). Have you tried using mpich (or some other mpi implementation) instead to see if that works? Also, have you tried using 1.10.0?

Scot


On Jul 23, 2016, at 1:12 PM, victor sv <[hidden email]> wrote:

Hi again,

In order to help to understand what is happening,  I've implemented and example using the HDF5 library.

You can see it in the following link:
https://github.com/victorsndvg/XH5For/blob/master/src/examples/hdf5_performance_test/ch_unstructured_hexahedron_perf.f90

I've compiled and launched it reproducing the previous mail and I'm getting the same behavior/errors with the new code.

I also compiled and linked against HDF5 1.8.17 by myself with the same results.

I'm not using any tunning hint ... this could be the problem?

Thanks!
Víctor



2016-07-19 11:08 GMT+02:00 victor sv <[hidden email]>:
Hi all,

I'm developing XH5For, an OO fortran light-weight XDMF/HDF5 layer, and now I would like to test its scalability, but I'm lost with an issue that i'm getting.

I'm going to try to explain it as best as I can.

I'm performing some weak scalability tests on Marenostrum III (GPFS file system) using Collective Writing with the Contiguous HyperSlab strategy. I'm running tests with 1, 16, 32, 64, 128, 256, 512, 1024 and 2048 MPI tasks.

All seems to work as expected except for the 2048 MPI tasks test, where I think I'm getting an MPI interlock (it's still running without doing anything till the job time exceeds and the job is killed).

After that, i try to reproduce an error with a number of MPI tasks between 1024-2048, and I could get the following error message while launching a smaller size job, with 1164 MPI tasks:

HDF5-DIAG: Error detected in HDF5 (1.8.16) MPI-process 1009:
  #000: H5Dio.c line 271 in H5Dwrite(): can't prepare for writing data
    major: Dataset
    minor: Write failed
  #001: H5Dio.c line 352 in H5D__pre_write(): can't write data
    major: Dataset
    minor: Write failed
  #002: H5Dio.c line 789 in H5D__write(): can't write data
    major: Dataset
    minor: Write failed
  #003: H5Dmpio.c line 529 in H5D__contig_collective_write(): couldn't finish shared collective MPI-IO
    major: Low-level I/O
    minor: Write failed

I have used the following libraries/versions during compilation stage:
Here you can see how I open the HDF5 file for Collective Writing:
https://github.com/victorsndvg/XH5For/blob/master/src/lib/hdf5_handler/hdf5_handler.f90#L531

Note: ENABLE_MPI, ENABLE_HDF5. ENABLE_PARALLEL_HDF5 definition flags are enabled

Could anyone give me some ligth about this?

I would greatly appreciate your help!

Thank you in advance,
Víctor.



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5

_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Hyperslab collective writing crash with HDF5 1.8.16

victorsv
Thank you Scot,

I've checked XH5For with, intel/16.0.1, openmpi/1.8.5 and HDF5/1.8.17 and it works with 2048 MPI tasks.
I'm launching more tests to check the performance.

Till the moment I'dont know how to handle CMake Fortran flags to suggest my preferences instead of force ... I will try to fix it.

I'm happy for receiving your support, Thank you so much again! :)

Best regards,
Víctor.


2016-07-26 21:57 GMT+02:00 Scot Breitenfeld <[hidden email]>:
Using hdf5 1.8.17, I ran your program on a lustre system using Cray’s mpi wrappers for the intel compiler on 3072, 1536, 1200, 2016 processes with no issues (it completed in under 20s). If you have not already done so, you might want to make sure  the hdf5 parallel test programs all passed.

FYI, I managed to get it to compile by adding the static flag for the compiler, I had to hard code the option in cmake. Here is what I used to compile it,

cmake \
    -D CMAKE_C_FLAGS:STRING="-g" \
    -D CMAKE_Fortran_FLAGS:STRING="-g -static" \
    -D CMAKE_BUILD_TYPE:STRING="Debug" \
    -D XH5For_ENABLE_HDF5=ON \
    -D XH5For_ENABLE_MPI=ON \
    -D XH5For_ENABLE_TESTS=ON \
    -D XH5For_ENABLE_EXAMPLES=ON \
    -D BUILD_SHARED_LIBS=OFF \
    -D HDF5_INCLUDE_DIRS:PATH=$HDF5_DIR/include \
    -D HDF5_LIBRARY_DIRS:PATH=$HDF5_DIR/lib \
    $XH5For
 
Scot


On Jul 26, 2016, at 11:09 AM, victor sv <[hidden email]> wrote:

Hi Scott,

thanks for your response.

I've tried with GNU compilers, OpenMPI and HDF5 (1.8.15, 1.8.16, 1.18.17), but in a local scope or small cluster, not in Marenostrum.

Ok, I'm going to try to compile all this tools to reproduce this environment and I will tell you if there are some news.

Can you report the compilation errors you are talking about? I can help you in this point or fix any issue!

Thank you again for your effort!
Víctor.

2016-07-26 16:41 GMT+02:00 Scot Breitenfeld <[hidden email]>:
I suspect that it is an issue with the impi implementation. I have not been able to compile your code on our systems due to various compilation errors (this is using a different compiler as I don’t have access to impi). Have you tried using mpich (or some other mpi implementation) instead to see if that works? Also, have you tried using 1.10.0?

Scot


On Jul 23, 2016, at 1:12 PM, victor sv <[hidden email]> wrote:

Hi again,

In order to help to understand what is happening,  I've implemented and example using the HDF5 library.

You can see it in the following link:
https://github.com/victorsndvg/XH5For/blob/master/src/examples/hdf5_performance_test/ch_unstructured_hexahedron_perf.f90

I've compiled and launched it reproducing the previous mail and I'm getting the same behavior/errors with the new code.

I also compiled and linked against HDF5 1.8.17 by myself with the same results.

I'm not using any tunning hint ... this could be the problem?

Thanks!
Víctor



2016-07-19 11:08 GMT+02:00 victor sv <[hidden email]>:
Hi all,

I'm developing XH5For, an OO fortran light-weight XDMF/HDF5 layer, and now I would like to test its scalability, but I'm lost with an issue that i'm getting.

I'm going to try to explain it as best as I can.

I'm performing some weak scalability tests on Marenostrum III (GPFS file system) using Collective Writing with the Contiguous HyperSlab strategy. I'm running tests with 1, 16, 32, 64, 128, 256, 512, 1024 and 2048 MPI tasks.

All seems to work as expected except for the 2048 MPI tasks test, where I think I'm getting an MPI interlock (it's still running without doing anything till the job time exceeds and the job is killed).

After that, i try to reproduce an error with a number of MPI tasks between 1024-2048, and I could get the following error message while launching a smaller size job, with 1164 MPI tasks:

HDF5-DIAG: Error detected in HDF5 (1.8.16) MPI-process 1009:
  #000: H5Dio.c line 271 in H5Dwrite(): can't prepare for writing data
    major: Dataset
    minor: Write failed
  #001: H5Dio.c line 352 in H5D__pre_write(): can't write data
    major: Dataset
    minor: Write failed
  #002: H5Dio.c line 789 in H5D__write(): can't write data
    major: Dataset
    minor: Write failed
  #003: H5Dmpio.c line 529 in H5D__contig_collective_write(): couldn't finish shared collective MPI-IO
    major: Low-level I/O
    minor: Write failed

I have used the following libraries/versions during compilation stage:
Here you can see how I open the HDF5 file for Collective Writing:
https://github.com/victorsndvg/XH5For/blob/master/src/lib/hdf5_handler/hdf5_handler.f90#L531

Note: ENABLE_MPI, ENABLE_HDF5. ENABLE_PARALLEL_HDF5 definition flags are enabled

Could anyone give me some ligth about this?

I would greatly appreciate your help!

Thank you in advance,
Víctor.



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5

_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Loading...