Throttled workflows push steps into the ready queue multiple times.#136
Merged
Throttled workflows push steps into the ready queue multiple times.#136
Conversation
cb4ecc5 to
c4b6a66
Compare
Collaborator
|
@FrankD412 -- I ran a few tests with and without the |
Member
Author
|
@jsemler -- It's good that didn't appear since it was the main symptom. If your tests indicate that a study runs to completion with and without throttling then this issue should be fixed. It appeared from the logs last week that steps were only run once now, so we should be good. |
jsemler
approved these changes
Aug 20, 2018
FrankD412
added a commit
that referenced
this pull request
Sep 30, 2018
* Update setup.py to 1.1.3dev * Updated the getting started documentation * Added examples for running both lulesh specs * Flux for Spectrum bugfixes (#116) * Additional arguments can be passed through batch. * Correction of the use of extend to append. * Removal of OMPI vars from env. * Reversal of env altering. * Addition of mpi exe to batch. * Removal of -gpu flag * Addition of "allocated" to PENDING set. * Correction of the check_status method call. * Correction of the Flux import * Addition of the EnvironmentError to try/catch for check_status. * Addition of jobid to submission INFO logging. * Workflow setup fix (#117) * Fixed a workflow setup issue caused by spaces in the spec name * Additional dirname formatting * Refactor of the Study class to breakdown complex APIs (#118) * Minor docstring correction. * Change to study setup API to break out workspace creation. * Update Maestro frontend to use new Study method. * Renamed Study.setup to be better communicate its functionality. * Correction to style for configure_study method. * Moved environment application to add_step. * Split out acquiring environment elements to its own method. * Updates to staging checks. * Renamed _setup_linear and _setup_parameterized from setup to stage for clarity. * Tweaks and addition of starting to SpectrumFluxAdapter. (#119) * Updated the LULESH examples to use the LULESH git repository for cloning the repo (#121) * Add the generation of metadata to Study construction. (#120) * Addition of method and tweaks to support metadata. * Addition of call to the method for generating metadata. * Correction to encode YAML string for py3 compatibility. * Another attempt at py2-py3 compatible encoding. * Addition of load_metadata method. * Additional debug logging for the Flux Spectrum Adapter (#122) * Additional logging for debugging of the adapter. * Additional logging. * Updated Exception.message references to Exception.args (#125) * Fix status.csv 'State' column writeout. (#127) * #107 Enhance/return codes (#128) * #107 Added StudyStatus enumeration Changed DAG execute_ready_steps to return StudyStatus enum instead of a boolean conductor and maestro also accept StudyStatus and convert to return value * #128 fixed issues found in pull request * #107 added inline comments per PC * #107 pulling out _check_study_completion to avoid initial implementaion copy and paste error * #107 fixed line length issue * Bugfix that would cause the ExecutionGraph to not update when cancelled. (#131) * Addition of custom parameter generation when running studies. (#129) * Command line parameter for custom ParameterGenerators. * Addition of sample custom parameter generation. * Correction to missing parameters in SIZE. * Addition of utility method for importing custom gen py files. * Correction of syntax. * Conversion of DOS newlines to Unix. * Addition of call to check for custom code. * Correction of a bad docstring. * Addition of parameter metadata to Maestro. * Addition of parameters to study metadata. * Addition of copying of the parameter file to study workspace. * Correction to a bug for attempting a restart when a restart isn't specified. (#133) * Initial fix for steps that cannot restart. * Comments to clarify logic. * More flexible ExecutionGraph description API and logging of description. (#137) * Tweaks to how the ExecutionGraph adds descriptions. Added logging. * Updated study to use the new API and some minor refactor. * Fixed #134 * Fix to Variable verification. (#138) Fixes #123 * Updates to Maestro documentation (#114) * Updates to the README. * Added initial getting started example documentation * Addition of Quick Start documentation. (#130) * Addition of Quick Start. * Addition of cancellation status example. * Corrections to quick start and fix for build errors. * Addition of a LULESH study specification with heavier comments. (#132) * Start of a LULESH spec that's commented. * Addition of explanation of step structure. * Additional comments about steps and dependencies. * Minor tweaks to comments. * Addition of parameter block comments. * Addition of reference to launcher_token YAML * Tweaks and edits based on feedback. * Throttled workflows push steps into the ready queue multiple times. (#136) * Moved 'mark_submitted' into ExecutionGraph 'execute' * Conversion of ready_steps deque to use step names. * Capping the available slots to the length of ready_steps. * Fixes Popen output to use universal newlines. (#143) * Addition of a shell util to centralize process creation settings. * Switch to start_process method. * Addition of docstring for start_process. * Update to fix universal newline from false to true. * Addition of a check for a list and a hard set of shell to False. * Made conductor launch command a string. * Fix behavior of start_process for localscriptadapter. The shell=False flag was not being passed, nor were the env or cwd arguments being used in the actual function. * Clean up of start_process * Cleans up GitDependency errors to not rely on return codes. (#144) * Addition of a ping method. * Tweaks to error logging. * Moved specific existence exception out of return code check. * Cleanup of some merge characters. * Tweak for connectivity checking. * Addition of info to the error message. * Removal of git return codes since they don't mean anything. * Update version number to 1.1.3.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Originally from #133 (comment):
@jsemler -- Alright, after a good bit of digging around, I think I've narrowed the bug down. I apologize if this explanation is somewhat convoluted, but I'll do what I can to make it clear.
In the
ExecutionGraphmethodexecute_ready_steps, there are a number of back end sets that track various aspects of execution, including failed steps, in progress steps, completed steps, dependency completions, and various other things (which for this discussion aren't important). The major thing that gets tracked are the ready steps which are in adequeobject. When a step is deemed to be ready, itsStepRecordobject is pushed to theready_stepsqueue. However, in theexecute_ready_stepsmethod, the following code currently exists (line 785 in executiongraph.py):I noticed that in the log you provided that steps that were repeated would always trigger the "All dependencies completed. Staging." log message and I never saw the passing message in the
else. That's likely because when I refactored and cleaned up the method last time, I made the transition to tracking records by their name instead of the object in order to clean up a lot of therecord.namereferencing I was doing. Somewhere along the way, it looks like I forgot to transition theready_stepsqueue to the same structure and so it goes to check if the key is in thedequeand never finds it since it's a queue ofStepRecordobjects.I don't expect the fix to this will be difficult, so I'll try and have it completed today. We can still go through it if you'd like.
Edit: As an extra note, I think this issue only appears when throttling because that is the only time you'll have steps queued that are still in the
INITIALIZEDstate. If not throttling, you'd never run into a duplicate in the queue since those steps would be in at leastPENDINGwhen submitted.