[petsc-users] TSAdjoint multilevel checkpointing running out of memory

[Zhang, Hong]

Anton,

There is a memory leak bug in TSTrajectory for this particular case. It has already been fixed. Can you update your PETSc and retry? Thank you for reporting the issue to us.

Best,
Hong (Mr.)

On Dec 10, 2020, at 5:19 PM, Anton Glazkov <anton.glazkov at chch.ox.ac.uk<mailto:anton.glazkov at chch.ox.ac.uk>> wrote:

Dear Matt and Hong,

Thank you for your quick replies!
In answer to your question Matt, the application fails in the same way as with 5 checkpoints. I don’t believe the RAM capacity to be a problem though because we are running this case on a cluster with 64GB RAM per node, and we anticipate 0.1GB storage requirements for the 4 checkpoints.
The case is being run in MPMD mode with the following command:

aprun -n 72 /work/e01/e01/chri4903/bin/cascade-ng/checkpoints_gradients ../data/nl_adj_0-chkpts.ini -adjoint -vr "/…/…/…/…/sims/aaX-testcases/10-RotorStator/test/data/ic_0_chkpts.h5:/0000000000/field" -vTarg "/…/…/…/…/sims/aaX-testcases/10-RotorStator/test/data/targ_0_chkpts.h5:/targ/field" -vMet "/…/…/…/…/sims/aaX-testcases/10-RotorStator/test/data/targ_0_chkpts.h5:/metric_diag/field" -ts_trajectory_dirname ./test_directory_0 -ts_trajectory_type memory -ts_trajectory_max_cps_ram 4 -ts_trajectory_max_cps_disk 5000 -ts_trajectory_monitor : -n 80 /…/…/…/…/bin/cascade-ng/checkpoints_gradients ../data/nl_adj_1-chkpts.ini -adjoint -vr "/…/…/…/…/sims/aaX-testcases/10-RotorStator/test/data/ic_1_chkpts.h5:/0000000000/field" -vTarg "/…/…/…/…/sims/aaX-testcases/10-RotorStator/test/data/targ_1_chkpts.h5:/targ/field" -vMet "/…/…/…/…/sims/aaX-testcases/10-RotorStator/test/data/targ_1_chkpts.h5:/metric_diag/field" -ts_trajectory_dirname ./test_directory_1 -ts_trajectory_type memory -ts_trajectory_max_cps_ram 4 -ts_trajectory_max_cps_disk 5000 -ts_trajectory_monitor > log.txt 2> error.txt

I have attached the log.txt and error.txt to this email so that you can have a look at these. It seems to look ok until the OOM killer kills the job.

Best wishes,
Anton

From: Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>>
Date: Wednesday, 9 December 2020 at 01:38
To: Zhang, Hong <hongzhang at anl.gov<mailto:hongzhang at anl.gov>>
Cc: Anton Glazkov <anton.glazkov at chch.ox.ac.uk<mailto:anton.glazkov at chch.ox.ac.uk>>, petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov> <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>>
Subject: Re: [petsc-users] TSAdjoint multilevel checkpointing running out of memory
On Tue, Dec 8, 2020 at 6:47 PM Zhang, Hong via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>> wrote:
Anton,

TSAdjoint should manage checkpointing automatically, and the number of checkpoints in RAM and disk should not exceed the user-specified values. Can you send us the output for -ts_trajectory_monitor in your case?

One other thing. It is always possible to miscalculate RAM a little. If you set it to 4 checkpoints, does it complete?

  Thanks,

      Matt

Hong (Mr.)


On Dec 8, 2020, at 3:37 PM, Anton Glazkov <anton.glazkov at chch.ox.ac.uk<mailto:anton.glazkov at chch.ox.ac.uk>> wrote:

Good evening,

I’m attempting to run a multi-level checkpointing code on a cluster (ie RAM+disk storage with –download-revolve as a configure option) with the options “-ts_trajectory_type memory -ts_trajectory_max_cps_ram 5 -ts_trajectory_max_cps_disk 5000”, for example. My question is, if I have 100,000 time points, for example, that need to be evaluated  during the forward and adjoint run, does TSAdjoint automatically optimize the checkpointing so that the number of checkpoints in RAM and disk do not exceed these values, or is one of the options ignored. I ask because I have a case that runs correctly with -ts_trajectory_type basic, but runs out of memory when attempting to fill the checkpoints in RAM when running the adjoint (I have verified that 5 checkpoints will actually fit into the available memory). This makes me think that maybe the -ts_trajectory_max_cps_ram 5 option is being ignored?

Best wishes,
Anton



--
What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/<http://www.cse.buffalo.edu/~knepley/>
<error.txt><log.txt>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20201221/f46b2ab9/attachment.html>