suter [Wed, 6 Nov 2013 17:39:48 +0000 (18:39 +0100)]
handle pending asynchronous requests on finalize (should be harmless in
most cases, everybody should has completed)(cherry picked from commit a09da6c17ad98c22da2c043958f60e0336e9942b)
Arnaud Giersch [Wed, 6 Nov 2013 06:47:24 +0000 (07:47 +0100)]
Misc cleanups.
* stat'ing the directory before mkdir is just wrong (race condition),
remove this test and let fopen report any error;
* use mode 0777 for mkdir, and let the user adjust his umask;
* use xbt_free() instead of free();
* remove useless tests before xbt_{dict,dynar}_free();
* and the usual reindent.
(cherry picked from commit 332285d51119f8f98270655bf88f485a9f914df1)
change tesh test to account for new structure in time independent traces
Not ideal, as tesh is a bit too strict to allow us to really compare the files
(cherry picked from commit bf31c01ca3d4167bd748658b02b5938231f23fc5)
Make time independent tracing output by default in one file per process
Structure is filename.txt : list of all the files to open
folder filename.txt_files : all files, with unique names (so that they won't be erased by a new run, only the main one will)
Adds an option tracing/smpi/format/ti_one_file to only use one subfile and put every event in it
For larger runs this will avoid opening too many handles.
(cherry picked from commit dc0e44715d73224df3cff50a9a0f42e73b1afc8c)
Arnaud Giersch [Wed, 6 Nov 2013 06:47:24 +0000 (07:47 +0100)]
Misc cleanups.
* stat'ing the directory before mkdir is just wrong (race condition),
remove this test and let fopen report any error;
* use mode 0777 for mkdir, and let the user adjust his umask;
* use xbt_free() instead of free();
* remove useless tests before xbt_{dict,dynar}_free();
* and the usual reindent.
I suppose this commit is related to Laurent Pouilloux's work?
I don't think it's really ready to be committed into SimGrid:
- please add some provenance information (version of the testbed
description, date of generation at least)
- please point to the tools used to generate the platform file
(or commit them to git), and some documentation about how
to re-do the process
- please use the <cluster> tag to avoid such a gigantic file
- please don't override the previous g5k.xml file, but rather keep
it (for historical / XP reproducibility purposes). maybe name
the new file g5k-11.2013.xml or sthing?
change tesh test to account for new structure in time independent traces
Not ideal, as tesh is a bit too strict to allow us to really compare the files
Make time independent tracing output by default in one file per process
Structure is filename.txt : list of all the files to open
folder filename.txt_files : all files, with unique names (so that they won't be erased by a new run, only the main one will)
Adds an option tracing/smpi/format/ti_one_file to only use one subfile and put every event in it
For larger runs this will avoid opening too many handles.
Arnaud Giersch [Wed, 30 Oct 2013 15:50:48 +0000 (16:50 +0100)]
Slightly change the sleep duration.
The goal is to have the same results on 32 and 64 bits, despite the different precision
for floating point operations. To achieve this, we choose a sleep_delay that is not a
multiple of the other delays of the algorithm. Hope that it will be sufficient.
Augustin Degomme [Wed, 30 Oct 2013 13:46:16 +0000 (14:46 +0100)]
Add an Inception tesh to test TI output.
It uses pingpong executable to generate a TI trace.
Then it replays this TI trace, generating a second one.
They are then compared one to each other (after sorting), and should be identical
Augustin Degomme [Wed, 30 Oct 2013 13:42:49 +0000 (14:42 +0100)]
add Time independent output trace format ( TI ) to output traces that can then be replayed
For now only handles SMPI
Using -trace-ti flag with smpirun should output a smpi_simgrid.txt file by default,
which could be then used for replay. All outputs are for now merged into only one file.
Augustin Degomme [Tue, 22 Oct 2013 15:35:01 +0000 (17:35 +0200)]
as replay uses a weird case for labelling its inputs, make it case insensitive
This will make TI traces generation easier for simgrid, without breaking existing replay