env_variables.txt

------------------------------------------------------------------------
Examples of optional environment variables - bash syntax
------------------------------------------------------------------------

export ADD_TIMESTAMP=yes   
Adds a timestamp to the file names of profiler outputs.
The default is to append a job id and the MPI rank to profiler outputs.

export COLLECTIVE_BARRIER=[yes|MPI_Bcast,MPI_Allreduce]
Adds an internal barrier to all collective routines, or to a specific list.
The profiler outputs will then show the measured synchronization time.

export ENABLE_TRACING=yes
Turns on event tracing in a region marked by MPI_Pcontrol(1); ...; MPI_Pcontrol(0);

export PROFILE_BEGIN_TIME=500; export PROFILE_END_TIME=520
Defines a time window for collection of profile data.
Units are seconds after the program calls MPI_Init().
When using time window profiling, the user should also set SAVE_LIST.

export PROFILE_BY_CALL_SITE=yes
This option tags each MPI call with its instruction address in the
program text section of the executable file.  The output will then
show MPI profile data for each distinct call site in the code.

export PROFILE_BY_COMMUNICATOR=yes
Separately reports MPI calls in buckets based on the size of the communicator.

export SAVE_ALL_TASKS=yes
Setting this to yes will produce one text output file for each MPI rank.

export SAVE_LIST=0,20,40
Use a comma separated list to specify the ranks that will produce profiler outputs.

export TRACEBACK_ABORT=yes
Produce a call-stack (instruction addresses) from MPI_Abort().

export TRACEBACK_ERRORS=yes
Produce a call-stack (instruction addresses) when any MPI function fails.

export TRACEBACK_LEVEL=2
Sets a level for how far up the call chain to start when reporting call stacks.

export TRACE_ALL_EVENTS=yes
Starts the collection of time-stamped event records in MPI_Init().

export TRACE_ALL_TASKS=yes
Requests time-stamped event collection from all MPI ranks.
The default is that records are saved for ranks 0-255.

export TRACE_BEGIN_TIME=200; export TRACE_END_TIME=210;
Defines a time window for collecting time-stamped event records.
Units are seconds after the application calls MPI_Init().

export TRACE_BUFFER_SIZE=48000000
Sets the size of the local buffer used by each MPI rank to save time-stamped
event records.  Units are in bytes, and each record takes 48 bytes.  The default
value is 4800000 bytes, which is enough for 10^5 events per rank.

export TRACE_DIR=/path/to/profiler/output
Optionally set this to send profiler output to a specific directory.
The default is to write outputs in the working directory for each rank.

export TRACE_DISABLE_LIST=MPI_Iprobe,MPI_Comm_rank
Optionally specify a list of MPI calls to exclude from event tracing.
The listed routines will still appear in the cumulative profile data.

export TRACE_MIN_RANK=40;  export TRACE_MAX_RANK=50;
Specify a contiguous band of MPI ranks that will save event records for trace
visualization.  The defaults are to save data for ranks 0-255.

export TRACE_SEND_PATTERN=yes
This turns on tracking of source and destination for MPI calls that have
specific destination ranks.  The accumulated data is reported in text files
that characterize the communication pattern : how many messages and bytes 
are sent from a given MPI rank to various destination ranks.

export TRANSLATE_RANKS=yes
When TRACE_SEND_PATTERN=yes (described above) you can optionally set this
variable to translate the ranks from an arbitrary communicator to the rank
in MPI_COMM_WORLD.  Such a translation may be required to properly characterize
the communication pattern.  The default is no, in which case the destination
rank that is reported is just the value that appeared in the MPI call.