June 2011 - Darshan-users - mailman3.cels.anl.gov

[Bug report] Darshan and Fortran (mpif90)
by 8schroed＠informatik.uni-hamburg.de 21 Jul '11

21 Jul '11

Hello, I tried to run Darshan with Fortran. For testing I wrote a short program which is using MPICH2 and MPI-I/O (see the attachment hello-world-mpi.f90). After compiling the code with Darshan and running the executable, Darshan didn't generate a logfile. To find the cause I analysed the Mapfile "$tmpfile" which is created with the build. The main interesting part in this file shows the symbols of MPI-functions: 0x0000000000400b58 mpi_comm_rank_ 0x0000000000400b68 mpi_file_close_ 0x0000000000400b78 mpi_file_open_ .... 0x0000000000400bd8 mpi_file_write_ 0x0000000000400be8 mpi_init_ 0x0000000000400bf8 mpi_finalize_ 0x0000000000400c08 mpi_comm_size_ But in mpif90 (of Darshan) the "grep"-command in line 84 is searching for "MPI" in upper case even though "MPI" is written in lower case in the Mapfile: grep MPI $tmpfile As a result the script "mpif90" of MPICH2 is executed instead of Darshan, because "grep" sent an exitcode equal 1. By adding the "-i" option, "grep" ignores case distinctions and the exit code is equal 0. Now mpif90 of Darshan will be used and after running the program there is a logfile in the log directory. Please take note of this bug and fix it in your Darshan-mpif90 to avoid the problem of case sensitivity. Best regards, Sandra

4 3

Darshan 2.1.1 release
by Phil Carns 23 Jun '11

23 Jun '11

Darshan version 2.1.1 is now available for download. It includes some performance improvements and bug fixes, as well as a new utility for converting log files. http://www.mcs.anl.gov/darshan/2011/06/23/darshan-2-1-1-release/ Changelog: * new darshan-convert command line utility for converting existing log files, with optional anonymization and optional bzip2 compression * bzip2 support in command line utilities (but not in the darshan library itself) * updated log file format that allows for string key/value pairs to be stored in the header * added ability to set MPI-IO hints when writing darshan log - at configure time: --with-log-hints - at run time: DARSHAN_LOGHINTS environment variable * bug fix contributed by Sandra Schrï¿½der: use case-insensitive search for MPI symbols in Fortran wrapper script * performance bug fix: remove unecessary call to MPI_File_set_size when writing log * added --with-logpath-by-env configure option to allow absolute log path to be specified via environment variable -Phil

1 0

Darshan Calibration
by ketan 07 Jun '11

07 Jun '11

Hello, Per my previous mails, I have been now successfully able to use Darshan for our programs. However, it seems that Darshan reports are not reflecting the amount of data produced by our program. More specifically, our program produces about 136K of data while Darshan plots seem to be reporting only about 1K. To combat this and to callibrate Darshan, we thought of profiling the UNIX coreutils dd program with Darshan. However, while trying to do this, when I insert the MPI_Init() and MPI_Finalize() in dd source, I get linker errors on compilation. So, my questions are: 1. Is there a definitive way of finding how Darshan intercepts the program? 2. Can I cross verify with some other tool what Darshan has produced is indeed consistent? 3. Are there any calibration parameters that I can set while using Darshan? 4. Has someone tried using Darshan with coreutils programs? Attaching two Darshan reports corresponding to our findings. Best Regards, Ketan

4 6