Add environment variable flag to python

[AdamSimpson]

Add flag to enable per process/color environment variables to be set. List of environment variables should be passed into WRAPRUN_FILE as fourth column in the format "key1=value2;key2=value2".

[avmarchenko]

Hello, was there ever any progress on this? I glanced at the code and it seems that (if using Python) when calling and instance of the Wraprun class's launch method there is an os.environ.update call (using the class instance's env (_env) - a dict of env var names/values I presume)? If I need to set a per task env variable can I accomplish it using Python or does launch set the global environment only? Can it be done via command line (without Python)? Could a hack be made to _subprocess_args to enable env defintions prior to aprun calls? Is this not possible because split.c doesn't know about env vars?

[AdamSimpson]

The underlying C library supports per-process environment variables although there is no easy way to set them currently. Paging @mpbelhorn to see what he thinks.

[mpbelhorn]

Per-process envvars can probably be done, but I don't think setting them in the Wraprun.launch environment within python is useful. That's just the environment in which aprun is called on the batch node. I think we'd need to write these variables to the WRAPRUN_FILE and parse/set them from the C library for each process.

[AdamSimpson]

The C library parses the fourth column in the WRAPRUN_FILE for environment variables. The column must be formated as key1=value2;key2=value2

[avmarchenko]

The python Wraprun class writes the WRAPRUN_FILE correct? Does that mean modifications only need to be made to the py side of things?

[mpbelhorn]

Yes, all the additions would need to be made in the python code. @AdamSimpson where are envvars currently added to the 4th WRAPRUN_FILE column? This looks like it was added in a very early commit. Is this vestigial code from when we were trying to pass all the rank parameters via envvars?

If that column is not already used for anything, then at a minimum, the wraprun.task.Rank object needs to be updated to have it's FILE_CONTENT tuple have an additional item for envvars that is by default an empty string. An argument would need to be added to wraprun.options.GroupOptions to parse input strings as Adam described, and the rank instantiation and ctor would need to add a keyword arg to set the envar string obtained from the option parser. I'm not immediately sure if any other changes would need to be made.

If the column is already used then we will need to make sure whatever is currently put there is preserved and merged with what wraprun.api._update_file will write there.

[AdamSimpson]

Ah good call, it's the forth column not the third. All of the logic appears to be present in libsplit to handle a fourth column of environment variable key-values if present.

[mpbelhorn]

Yeah, I'll start developing this feature. I don't think it'll take too long to implement.

[mpbelhorn]

Ok, I haven't tested it yet but the commit I just pushed appears to produce the correct form for the WRAPRUN_FILE lines:

$ export WRAPRUN_PRELOAD=none
$ wraprun --w-debug -n 3 --w-env 'key1=herp;key2=derp,key3=perp,,key4=foo' exe
BEGIN WRAPRUN DEBUGGING INFO
 Aprun call signature:
   aprun -b -n 12 exe

 Environment variables:
   LD_PRELOAD=none
   WRAPRUN_FILE=/autofs/nccs-svm1_home2/belhorn/development/wraprun/python/wraprun_H1PwrC.tmp
   W_IGNORE_SEGV=1
   W_REDIRECT_OUTERR=1
   W_UNSET_PRELOAD=1

 Internal state:
   Wraprun(ranks=12, colors=4)

 Tempfile contents:
   00|0 ./ unnamed.31601_w15.0 key1=herp;key2=derp
   01|0 ./ unnamed.31601_w15.0 key1=herp;key2=derp
   02|0 ./ unnamed.31601_w15.0 key1=herp;key2=derp
   03|1 ./ unnamed.31601_w15.1 key3=perp
   04|1 ./ unnamed.31601_w15.1 key3=perp
   05|1 ./ unnamed.31601_w15.1 key3=perp
   06|2 ./ unnamed.31601_w15.2 
   07|2 ./ unnamed.31601_w15.2 
   08|2 ./ unnamed.31601_w15.2 
   09|3 ./ unnamed.31601_w15.3 key4=foo
   10|3 ./ unnamed.31601_w15.3 key4=foo
   11|3 ./ unnamed.31601_w15.3 key4=foo
END WRAPRUN DEBUGGING INFO

Just need to deploy this on Titan and run the tests. I pushed it to the development branch so I can write a spack package to install it for testing.