Output a report about tasks that use more than 5% of the runtime in a step

For inspiration.

added 1 commit

• 36344f71 - Added a user-defined threshold to trigger a dump of the task graph

Compare with previous version

added 343 commits

• 36344f71...e6590cea - 341 commits from branch master
• ae845b37 - Output a report about tasks that use more than 5% of the runtime in a step
• 8157c486 - Added a user-defined threshold to trigger a dump of the task graph

Compare with previous version

I have used your initial approach to actually trigger a dump of the full task data. The threshold is user defined.

added 1 commit

• 0d61877e - Also update the call to the stats dump done when repartitioning

Compare with previous version

Can we extend the warnings at:

https://gitlab.cosma.dur.ac.uk/swift/swiftsim/blob/longtask-report/examples/main.c#L399

to make it clear that we need to enable task debugging and also require -y, for this option to work.

added 1 commit

• 3b4fe407 - Make the code abort if --dump-tasks-threshold is used without…

Compare with previous version

I have added a check forcing the configuration of SWIFT_DEBUG_TASKS to use the new feature.

Thanks, nearly ready, can we add this new command-line option to the various docs:

• README
• README.md
• doc/RTD/source/CommandLineOptions/index.rst

think that's it.

Thanks, I'll add these.

The other thing I am wondering about is whether this is the behaviour I want when running over MPI. Do I want the information only from the node that had the long task or from all nodes? The former is probably fine for now.

Well you get them all at the moment. Could be interesting from a balance point of view. A simpler dump would be just a cut-n-paste of the MPI version.

If I run with --dump-tasks-threshold but without -y at the moment, I'll hang over MPI on the first step where only 1 rank has a task above the threshold.

added 1 commit

• a04b39d4 - Update the documentation with the new runtime parameter

Compare with previous version

added 1 commit

• 9a0b874b - Prevent the use of the new analysis feature over MPI for now.

Compare with previous version

I am now preventing users from using this over MPI. If this proves useful, I'll extend this to the multi-node version.

I see, so at some point you'll need to reduce this over all the nodes, but it can wait. I'll accept for now.

OK for you?

Yes, thanks!

merged

mentioned in commit b4e619f7