topic OpenMP related I/O error in Intel® Fortran Compiler

OpenMP related I/O error

thornpig — Tue, 12 Oct 2010 21:39:27 GMT

I am running a numerical model in "dm+sm" parallel mode on a cluster with 64 bit linux OS which has more than 2 nodes and 16 cpus per node. The compilation of the model with "dm+sm" seemed successful.

When I set OMP_NUM_THREADS to 2 or greater , the model was terminated with an error message :

forrtl: severe (40): recursive I/O operation, unit 0, file unknown

When I set OMP_NUM_THREADS to 1, which acctually disabled the openmp function, the model ran successfully.

Please help me out. Thanks a lot !

Below is my PBS job file:

#!/bin/csh

#PBS -l nodes=2:ppn=8

#PBS -m ae

setenv OMP_NUM_THREADS 2

time mpirun wrf.csh

where "Wrf.csh" unlimits the stacksize and executes the model as follows:

#!/bin/csh

limit stacksize unlimited

exec wrf.exe

OpenMP related I/O error

jimdempseyatthecove — Wed, 13 Oct 2010 17:18:47 GMT

Is your file I/O context held in Thread Local Storage?

Locate in your code where (you suspect) the erronious I/O statement is located. Insert some diagnostic code

!$OMP CRITICAL
WRITE(*,*) 'Debug IO ', omp_get_thread_num()
bSomeError = .false.
(your I/O statement(s) here)
goto 12345
(your I/O error code here)
bSomeError = .true.
12345 continue
!$OMP END CRITICAL
if(bSomeError) goto (your error lable here)

Jim

OpenMP related I/O error

Martyn_C_Intel — Fri, 15 Oct 2010 22:46:48 GMT

The "recursion" in thisparticular error message sometimes means that an error occurred during a write to stdout/stderr, and so when the RTL tries to write a diagnostic to the same unit, it is considerd as a recursive use of that unit. But it might indicate an error or thread safety issue on some other unit.

Along the same lines as Jim's suggestion, you should make sure that you are linking to the threadsafe version of the Fortran RTL. This should happen automatically if you use the compiler driver to link and have -openmp on the link line. But if you link to the OpenMP library explicitly, then you may need to put -threads on the link line, or link to libifcoremt instead of libifcore.

OpenMP related I/O error

thornpig — Wed, 20 Oct 2010 15:52:38 GMT

Thank you Jim and Martyn!

I am trying to figure out your suggestions and will give you feedback later.

OpenMP related I/O error

Yongjia_S_ — Wed, 27 Oct 2010 23:20:06 GMT

I gotsimilarproblem to run WRF model compiled by intel compiler with openmp. wrf.exe can run with omp_num_threads=1, but get same error when omp_num_threads>1. However, when I use gfortran/gcc to compile this model with openmp, it can run successfully when omp_num_threads>1. any clue? thanks.

OpenMP related I/O error

jimdempseyatthecove — Thu, 28 Oct 2010 04:21:37 GMT

One other thing to look at is:

Does your I/O statement(s) within the parallel region(s) attept to transfer outside the parallel region (on error, eof, ...)?

It is not valid to do so. (place these branch targets inside the parallel region)
The compiler should warn of this coding error.

Jim Dempsey

OpenMP related I/O error

Yongjia_S_ — Thu, 28 Oct 2010 20:18:02 GMT

I don't undersand the mean in Jim's code. what is "(your I/O statement(s) here)" and "(your I/O error code here)", are they same io sentence I suspect in my program? what is "(your error lable here)"? thanks.

OpenMP related I/O error

jimdempseyatthecove — Thu, 28 Oct 2010 20:36:40 GMT

Quoting yongjia.song@gmail.com

!$OMP PARALLEL
...
READ(YourInUnit, 100, err=999, end=888) arglist
100 FORMAT(...)
...
GOTO 12345
! error label
! *** MUST reside within sameparallel region of READ(..., err=label)
999ErrorFlag = .true.
GOTO 12345 ! to label at end of parallel region
! *** MUST reside within sameparallel region of READ(..., end=label)
888 EndFlag = .true.
GOTO 12345
...
12345 CONTINUE
!$OMP END PARALLEL
if(ErrorFlag) GOTO 9999
if(EndFlag) GOTO 8888
...

The labels 999 and 888 are dispatched to froma READ statementwithin the parallel region and therefore must reside within the same parallel region. Following the exit of the parallel region, test for end and/or error conditions.

"A structured block of code is a collection of one or more executable statements with a single point of entry at the top and a single point of exit at the bottom."

I hope this clears up the issue.

Jim Dempsey

OpenMP related I/O error

Yongjia_S_ — Thu, 28 Oct 2010 22:39:47 GMT

Thanks, Jim, I understand now.