Skip to end of metadata
Go to start of metadata

This page has a list of known issues for the Intel compilers (all versions).

 

Use of MKL library can cause irreproducible results

OpenIFS includes a compilation configuration for the Intel compiler with the Intel MKL library (for optimized LAPACK/BLAS). However, please be aware use of this library can cause the model to be irreproducible, even on the same core count in successive runs. We recommend not using it if reproducibility is a concern.

OpenIFS also only provides a compilation configuration for the MKL and the Intel library. Linking MKL with other compilers is possible, though complicated and is not tried or tested with OpenIFS.

For help with linking the MKL library with other compilers, please see: https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor

 

 

OpenIFS can fail with Intel compiler at -O2

There is an issue with OpenIFS when compiling with the Intel compiler at optimization level -O2 or above on chipsets that support SSE4.1 & AVX instructions. Intel compilers are generally more aggressive at optimisations for -O2 than other compilers.

Users will see failure with the T21 test job similar to the following:

Sample failure message
signal_harakiri(SIGALRM=14): New handler installed at 0x432110; old preserved at 0x0
 ***Received signal = 8 and ActivatED SIGALRM=14 and calling alarm(10), time = 6.18
 myproc#1,tid#1,pid#27600,signal#8(SIGFPE): Received signal :: 123MB (heap), 125MB (rss), 0MB (stack), 0 (paging), nsigs 1, time 6.18
 tid#1 starting drhook traceback, time = 6.18
 myproc#1,tid#1,pid#27600: MASTER 
 myproc#1,tid#1,pid#27600: CNT0<1> 
 myproc#1,tid#1,pid#27600: CNT1 
 myproc#1,tid#1,pid#27600: CNT2 
 myproc#1,tid#1,pid#27600: CNT3 
 myproc#1,tid#1,pid#27600: CNT4 
 myproc#1,tid#1,pid#27600: STEPO 
 myproc#1,tid#1,pid#27600: SCAN2H 
 myproc#1,tid#1,pid#27600: SCAN2M 
 myproc#1,tid#1,pid#27600: GP_MODEL 
 myproc#1,tid#1,pid#27600: EC_PHYS_DRV 
 myproc#1,tid#1,pid#27600: >OMP-PHYSICS CLDPP T/S (1002) 
 myproc#1,tid#1,pid#27600: EC_PHYS 
 myproc#1,tid#1,pid#27600: CALLPAR 
 myproc#1,tid#1,pid#27600: SLTEND 

It arises because this compiler makes use of 2-way vectorization when compiling both branches of IF statements which can generate floating point exceptions if a zero divide is possible in the unexecuted branch and the IFS internal signal handler (DRHOOK) is enabled.

There are several possible workarounds:

  1. Compile the routines that cause the problem with lower optimisation, -O1. The routines affected are: sltend.F90, vsurf_mod.F90, vdfmain.F90, vdfhghtn.F90.
  2. Run with the environment variable: DR_HOOK_IGNORE_SIGNALS=8 to disable trapping of floating point exception signals (SIGFPE) by the model. This is not ideal as it will not catch other causes of floating point exceptions.
  3. Edit the code and insert the line:

     !DEC$ OPTIMIZE:1

    directly after the SUBROUTINE statement into the routines: sltend.F90, vsurf_mod.F90, vdfmain.F90, vdfhghtn.F90.

  4. Edit the intel-*.cfg configuration files in make/cfg and add lines to change the compile options specifically for these files.

OpenIFS uses a default of -O1 in the configuration files. If you increase the optimisation level, please be aware of this issue.

For more help with this issue, please contact openifs-support@ecmwf.int.

 

 

OpenIFS fails writing GRIB if grib_api compiled with Intel and -O2

We are aware of a problem in grib_api when using the Intel compiler that seems to affect different versions of grib_api and causes the model to fail with a floating point exception (SIGFPE). This is known to happen in the routine PRESET_GRIB_TEMPLATE or in the GRIB_F_SET_REAL8_ARRAY in the grib_api library. The advice is to reduce the optimization level when compiling grib_api to -O1 rather than -O2 or try a more recent version of the Intel compiler.

The error message that typifies this problem is:

OpenIFS log file
***Received signal = 8 and ActivatED SIGALRM=14 and calling alarm(10), time =    3.10
JSETSIG: sl->active = 0
signal_harakiri(SIGALRM=14): New handler installed at 0xabfa00; old preserved at 0x0
***Received signal = 8 and ActivatED SIGALRM=14 and calling alarm(10), time =    3.10
[myproc#1,tid#1,pid#14063]:  MASTER
[myproc#1,tid#1,pid#14063]:   CNT0<1>
[myproc#1,tid#1,pid#14063]:    SU0YOMB
[myproc#1,tid#1,pid#14063]:     SU_GRIB_API
[myproc#1,tid#1,pid#14063]:      PRESET_GRIB_TEMPLATE
JSETSIG: sl->active = 0
signal_harakiri(SIGALRM=14): New handler installed at 0xabfa00; old preserved at 0x0

or a traceback like this:

Traceback
[gdb__sigdump] : Received signal#8(SIGFPE), pid=-1
[LinuxTraceBack]: Backtrace(s) for program 'oifs38r1/make/intel_mkl-opt_conv/oifs/bin/master.exe' (pid=38451) :
(pid=38451): oifs38r1/src_conv/ifsaux/utilities/linuxtrbk.c:109  :  master.exe() [0xc14a2d]
(pid=38451):      oifs38r1/src_conv/ifsaux/support/drhook.c:884  :  master.exe() [0xac8ddb]
(pid=38451):                                          <Unknown>  :  libpthread.so.0(+0xf7e0) [0x7f59e215b7e0]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(log.L+0x23c) [0x7f59e60db98c]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xa7de4) [0x7f59e60a7de4]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0x9c6d4) [0x7f59e609c6d4]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(grib_pack_double+0x18) [0x7f59e6079847]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xc4814) [0x7f59e60c4814]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xc4890) [0x7f59e60c4890]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(grib_set_double_array_internal+0x68) [0x7f59e60c4921]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xa3a4a) [0x7f59e60a3a4a]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(grib_pack_double+0x18) [0x7f59e6079847]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xc4814) [0x7f59e60c4814]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xc4890) [0x7f59e60c4890]
(pid=38451):                                          <Unknown>  :  libgrib_api.so.0(+0xc4b3f) [0x7f59e60c4b3f]
(pid=38451):                                          <Unknown>  :  libgrib_api_f90.so.0(grib_f_set_real8_array_+0x51) [0x7f59e6380aea]
(pid=38451):                                          <Unknown>  :  libgrib_api_f90.so.0(grib_api_mp_grib_set_real8_array_+0x8a) [0x7f59e63858af]
(pid=38451): oifs38r1/src_conv/ifsaux/module/grib_api_interface.F90:358  :  master.exe() [0xb03bbd]

Note that the grib packing can also fail if the model has produced fields with a very large range of values, such that the grib library can't pack the values into a smaller bit range. For further help, please contact openifs-support@ecmwf.int.

 

 


2 Comments

  1. I think I am encountering the irreproducibility issue. I have compiled with

    export OIFS_COMP=intel_mkl

    export OIFS_BUILD=opt.

    I have only seen this twice, both in long term (10 year) aqua-planet simulations. I have ran lots of weather forecast type of simulations and have not encountered this with the same compilation. The first time I saw this irreproducibility behaviour the model climate was statistically the same but the run as not identical and therefore I wasn't too concerned. However the second time, when I re-ran the simulation the model climate is completely different - the Hadley cell moves polewards and significantly strengthens! I am fairly sure I haven't changed anything else but is this possible? What sort of irreproducibility issues have you encountered before? I think the long term solution is to re-compile with a different compiler option. My only other option on the linux cluster I am running on is GNU (gcc / gfortran). Are these a safer option?

    Victoria



    1. Hi Victoria,

      the issue with the MKL library if I remember correctly is that it may not always use memory consistently. By that I mean, when using vector instructions, depending on the memory alignment, it may split the vectorizable statements/arrays differently causing potential for rounding differences. This behaviour was seen in IFS some years ago, I have not tested it recently. Safe compiler options that _should_ reproducibility for intel is:  -fpe0 -fp-model precise -fp-speculation=safe. You could also try using the lapack code supplied with OpenIFS rather than the system one. To do this, set the environment variable OIFS_LAPACK=openifs. I have not tested this for reproducibility though. 

      Glenn