Project

General

Profile

Bug #13749

Data prep service causing jobs to exceed memory usage

Added by Thomas Warburton about 4 years ago. Updated almost 4 years ago.

Status:
Closed
Priority:
Normal
Assignee:
Tingjun Yang
Start date:
09/01/2016
Due date:
% Done:

0%

Estimated time:
Duration:

Description

When running the reconstruction for the 10 kt detector and the 35 ton CRY milliblock samples the jobs use over 2 GB memory. Looking at the memory tracker information for the jobs it appears that this is caused by the data prep service.

I attach the file paths to two files which I have used to verify this:
/dune/data2/users/warburton/DataPrep/CorrectedDetsim_countermu_20160902T002056_reco_milliblock.root
/pnfs/dune/scratch/dunepro/v06_04_01/detsim/prodMUSUN_dune10kt/10115587_48/prodMUSUN_dune10kt_28_20160826T220116_gen_c3882a40-e647-4aee-b6c1-d01b55ced06b_g4_detsim.root

In this directory there is also some piped output from me running the jobs where you can see the memory usage from each job.
/dune/data2/users/warburton/DataPrep/

The fcl files which I used were the standard reconstruction fcl files:
standard_reco_dune35tsim_milliblock.fcl
standard_reco_dune10kt.fcl

David, could you take a look at this please?

History

#1 Updated by David Adams about 4 years ago

I will have a look but it may not be before next week.

You might try running with only the data prep producer.

#2 Updated by Thomas Warburton about 4 years ago

Hi David,

We cannot finish MCC7.1 without fixing this.

I just ran the data prep separately and it appears to see reduced memory usage (for the milliblock sample at least), which is good. However when I then run the event display I am seeing that the pedestals have been incorrectly applied.
Looking at the output of ART_DEBUG_CONFIG for standard_detsim_dune35t.fcl and standard_detsim_dune35t_milliblock there are a number of differences:
  • The pedestal for the milliblock sample does not use fixed pedestals, but the data pedestals.
  • The AdcSuppress service is different and there is also an extra service called AdcSuppressonService (with typo)
  • There are also extra services such as AdcDeconvolutionService

The first two problems go away if I add the fcl overwrites at the bottom of standard_detsim_dune35t.fcl to standard_detsim_dune35t_milliblock.fcl. I think this means that they are getting overwritten when we update the services to be used for the milliblock sample.

When I then run the event display on this file I do not see errors with the pedestals in the raw or reconstructed views.
I do see strange black bars in the reconstructed tab for TPCs 4 and 5....
Is it ok for me to push these changes?

#3 Updated by David Adams about 4 years ago

Karl:

For validation of the new 35t data prep, please see #12701. I did check that the new data prep gave the same results as the old caldata but maybe the old milliblock reco also had problems?

I am not sure what you propose to do. The overrides in standard_detsim_dune35t.fcl only affect producers that run after data prep. Here is the full file:

#include "standard_reco_dune35tsim.fcl" 

physics.producers.dbclusterdc.HitsModuleLabel:         "lineclusterdc" 
physics.producers.costrkdc.ClusterModuleLabel:         "lineclusterdc" 

physics.producers.dbcluster.HitsModuleLabel:           "hit35tcc" 
physics.producers.costrk.ClusterModuleLabel:           "linecluster" 

#Pandora configurations
physics.producers.pandoradc.HitFinderModuleLabel:           "dcheatcc" 
physics.producers.pandora.HitFinderModuleLabel:             "linecluster" 

physics.producers.linecluster.ClusterCrawlerAlg.TimeDelta:    [2,3,3]
physics.producers.lineclusterdc.ClusterCrawlerAlg.TimeDelta:    [2,3,3]

Changing these is fine with me but should not affect the output of data prep.

Am I missing something?

da

#4 Updated by Thomas Warburton about 4 years ago

Potentially, though I was not aware of any. Regardless pushing this fix will rectify it.

You show the fcl file for standard_reco_dune35tsim_cc.fcl. I was referring to standard_detsim_dune35t.fcl namely the following lines;
  1. Use fixed values instead of DB for pedestals.
    services.DetPedestalService: @local::dune_fixedpeds
  1. DetSim flags.
    physics.producers.daq.NoiseOn: true
    physics.producers.daq.PedestalOn: true
    physics.producers.daq.DistortOn: false
    physics.producers.daq.SuppressOn: true
  1. DetSim services.
    services.SimChannelExtractService: @local::scxgeneric
    services.ChannelNoiseService: @local::chnoiseold
    services.PedestalAdditionService: @local::padprovided
    services.AdcDistortService: @local::stuckbits
    services.AdcSuppressService: @local::zslegacy
    services.AdcCompressService: @local::cmpblock
  1. Disable bad channels.
    #services.IChannelStatusService.BadChannels: [ ]

I will push these changes to standard_detsim_dune35t_milliblock.fcl.

However, this is a separate point to data prep exceeding memory usage which still exists, most notably in the far detector geometry where running data prep separately did not reduce memory usage.

#5 Updated by David Adams about 4 years ago

Karl:

Sorry to be slow in responding. It seems there may be multiple issues being discussed here. It is easier for me if we open separate tickets in such cases.

Let's first settle the problems with detsim which may or may not be the source of the memory increase that led to this ticket. If I go back one release to v06_04_01, I see this for standard_detsim_dune35t_milliblock.fcl:

#include "standard_detsim_dune35t.fcl" 

services.DetectorPropertiesService.NumberTimeSamples:  32000
services.DetectorPropertiesService.ReadOutWindowSize:  32000
outputs.out1.fileName: "%ifb_%tc_detsim_milliblock.root"

I.e. milliblock uses the same configuration as the standard 35 ton except for number of samples. This should have been validated when I changed the standard 35-ton detsim. I gave a validation talk for the latter at the April 11 meeting of the FD sim/reco group but I don't see any evidence there that I validated milliblock production.

Did you find that this milliblock fcl did not work? Did you report that in Redmine or to me? Evidently you or someone found problems because the fcl was changed in v06_05_00 and is changed again in the head as you describe here. The new fcl duplicates a lot of code that is in standard_detsim_dune35t.fcl. This is probably not desirable as we would typically want future changes made in one to also appear in the other.

Thanks in advance for helping me to understand what is going on here. And if all this has been discussed in another ticket that has slipped my mind, please give me its ID and accept my apologies.

Thanks.
David

#6 Updated by Thomas Warburton about 4 years ago

David:

I would prefer one ticket, even if we just go through issues sequentially.

Tingjun and I decided ( privately ) that the milliblock sample should be changed to 15,000 ticks to better represent the data. This was the goal of the milliblock sample in the first place where we expected events which we 10 drift windows long ( at 500 V cm-1 ).

To achieve this I changed the parameters in services_dune.fcl but noticed that the change in the number of samples was not being passed to the sample correctly. This is because the NumberTimeSamples was hardcoded in this fcl file, so it was not inheriting from dune35t_milliblock_detsim_services. This was the reason for the first change made on 1st September.

When I then ran the detsim stage I noticed that many of the fcl parameters were now not set correctly, and so copied in the overrides from standard_dune35t_detsim.fcl to fix this, as you not above. This was the reason for the second commit made on 2nd September.
Looking at it with fresh eyes, the physics.producers.daq.XXXX: YYY lines are probably not needed as these will not be overwritten.
However the services.XXX lines are required as they are not set correctly in dune35t_milliblock_detsim_services or dune35t_detsim_services hence the need to overwrite them in standard_dune35t_detsim.fcl and here.

The third recent commit also made on 2nd September was because I noticed I was using dune35t_milliblock_reco_services not dune35t_milliblock_detsim_services.

I hope that helps to explain the commits which I made.
Karl

#7 Updated by David Adams about 4 years ago

So this should work?:

#include "standard_detsim_dune35t.fcl" 

services.DetectorPropertiesService.NumberTimeSamples:  15000
services.DetectorPropertiesService.ReadOutWindowSize:  15000
outputs.out1.fileName: "%ifb_%tc_detsim_milliblock.root"

Thanks.

da

#8 Updated by Thomas Warburton about 4 years ago

David:

Yes, that would work. However that seems even less optimal to me as the user is now not using dune35t_milliblock_detsim_services but is instead using dune35t_detsim_services and so the user has to know every difference between the two services. Luckily in this case the only difference is changing the number of the readout samples but there is no reason to think that this would always be the case.

Is there a good reason why the service overrides cannot go in services_dune.fcl? This is surely the logical place for them to go and not in the standard_detsim_dune35t fcl file.

Karl

#9 Updated by David Adams about 4 years ago

Karl:

I agree it is find to have and use dune35t_milliblock_detsim_services. Or to do as I suggest which I think is what we had before the change in the detsim service.

I just think the current top-level fcl is rather convoluted and difficult to understand.

Thanks.

da

#10 Updated by Thomas Warburton about 4 years ago

Ok, so we can leave standard_detsim_dune35t_milliblock.fcl as it currently is in the head of dunetpc?

Going back to the memory issue at hand.
  • There still remains a problem with memory usage when running dataprep as a standalone process on the full far detector geometry.
  • When running data prep by itself on the 35 ton milliblock sample the memory usage is acceptable ( at least for a 10 event sample ), when running it as part of standard_reco_dune35tsim_milliblock.fcl the memory usage sneaks over 2 GB ( for 10 events ) so I think in a full sample this would still cause problems.

Karl

#11 Updated by David Adams about 4 years ago

  • Status changed from New to Assigned

Karl:

The current version (option 1) of standard_detsim_dune35t_milliblock.fcl has this:

#include "standard_reco_dune35tsim.fcl" 

includedServices: @local::services

services: {
    @table::includedServices
    @table::dune35t_milliblock_reco_services
}

includedServices: @erase

outputs.out1.fileName: "%ifb_%tc_reco_milliblock.root"

So in addition to copying the producer chain from standard_reco_dune35tsim, it copies all the services and then overwrites those defined in dune35t_milliblock_reco_services. I suppose this is OK until we decide there is a service in standard_reco_dune35tsim that we would like to remove rather than overwrite.

An alternative (option 2) is:

#include "standard_reco_dune35tsim.fcl" 

services: {
  # Load the service that manages root files for histograms.
  TFileService: { fileName: "reco_hist.root" }
  TimeTracker:       {}
  MemoryTracker:     {}
  RandomNumberGenerator: {} #ART native random number generator
  message:      @local::dune_message_services_prod_debug
  FileCatalogMetadata:  @local::art_file_catalog_mc
  @table::dune35t_milliblock_reco_services
}

outputs.out1.fileName: "%ifb_%tc_reco_milliblock.root"

We no longer inherit any services from standard_reco_dune35tsim but I have verified this is equivalent to the current version. I prefer this because if someone adds something to standard reco it is not automatically added to milliblock reco. If they want it in both, they would add it to the shared block in services_dune.

Perhaps even better (option 3) would be to add include the explicit services above

  # Load the service that manages root files for histograms.
  TFileService: { fileName: "reco_hist.root" }
  TimeTracker:       {}
  MemoryTracker:     {}
  RandomNumberGenerator: {} #ART native random number generator
  message:      @local::dune_message_services_prod_debug
  FileCatalogMetadata:  @local::art_file_catalog_mc

in dune35tsim_reco_services and dune35t_milliblock_reco_services (e.g. by including them in shared definition). Then standard_detsim_dune35t_milliblock.fcl could simply be
#include "standard_reco_dune35tsim.fcl" 

services: @local::dune35t_milliblock_reco_services

Karl and Tingjun, please let me know which of these options you prefer. In the meantime, I will start to look at the memory issue.

#12 Updated by David Adams about 4 years ago

  • Assignee changed from David Adams to Thomas Warburton

I processed 10 cosmic events with full milliblock reco (standard_reco_dune35tsim_milliblock.fcl) and see this memory usage:

Events with large Vsize (Mbytes)        Vsize      Δ Vsize      RSS         Δ RSS
====================================================================================
  run: 10000002 subRun: 0 event: 4    1793.836        0        941.609        0
  run: 10000002 subRun: 0 event: 5    1793.836        0        941.629      0.020
  run: 10000002 subRun: 0 event: 6    1793.836        0        941.629        0
  run: 10000002 subRun: 0 event: 7    1793.836        0        941.637      0.008

I then split the processing to first run only dataprep:
Events with large Vsize (Mbytes)        Vsize      Δ Vsize      RSS         Δ RSS
====================================================================================
  run: 10000002 subRun: 0 event: 5    1483.090     181.328     564.277     117.523
  run: 10000002 subRun: 0 event: 1    1438.273     348.156     522.445     285.848
  run: 10000002 subRun: 0 event: 2    1301.762        0        445.070      0.449
  run: 10000002 subRun: 0 event: 4    1301.762        0        446.754        0

and then the rest of the reco chain (no dataprep):
Events with large Vsize (Mbytes)        Vsize      Δ Vsize      RSS         Δ RSS
====================================================================================
  run: 10000002 subRun: 0 event: 3    1778.629     125.422     925.395     124.750
  run: 10000002 subRun: 0 event: 5    1774.633     30.598      922.152     30.613
  run: 10000002 subRun: 0 event: 8    1765.668     71.762      913.203     71.773
  run: 10000002 subRun: 0 event: 4    1744.035     -35.078     891.539     -34.785

Comparing these, I conclude the dataprep adds very little to the memory usage.

Karl, you report we are using more memory than previous production/releases. I suppose this is due to one (or more) of the following:
1. More reco producers are being run
2. One of the reco producers has been modified and uses more memory
3. The upstream generation or simulation is allowing more data through

Of these, only the last can be due to the new dataprep (or detsim). Do you see that the detsim event size is larger now than it was in earlier releases?

For point 1, has the list of producers changed? If so, can you try running with the old list and see how that affects the memory usage?

I will wait for feedback before doing more here.

Thank you.

#13 Updated by Thomas Warburton about 4 years ago

  • Assignee changed from Thomas Warburton to David Adams

Hi David,

I will look at this further next week. I think that the milliblock sample may now be ok. However I believe that a problem still exists with the far detector geometry. Could you please have a look at this?

I noted in comment #4
However, this is a separate point to data prep exceeding memory usage which still exists, most notably in the far detector geometry where running data prep separately did not reduce memory usage.

Cheers,
Karl

#14 Updated by David Adams about 4 years ago

  • Assignee changed from David Adams to Thomas Warburton

Karl:

Is this only in full FD or in the 1x2x6 subdetector. Is the problem evident with single muons?

I will repeat the above study with this.

#15 Updated by Thomas Warburton about 4 years ago

I did not notice a problem when submitting the 1x2x6 geometries. It was with the MUSUN sample which is simulated in the full FD. The MUSUN sample simulates single muons.

You can find a g4 file to test at this location:
/pnfs/dune/scratch/dunepro/v06_04_01/g4/prodMUSUN_dune10kt/10109056_62/prodMUSUN_dune10kt_70_20160826T220124_gen_87e8d4ca-20f6-420f-b959-b366c3f8005d_g4.root
You can find a detsim file to test at this location:
/pnfs/dune/scratch/dunepro/v06_04_01/detsim/prodMUSUN_dune10kt/10115587_8/prodMUSUN_dune10kt_70_20160826T220124_gen_87e8d4ca-20f6-420f-b959-b366c3f8005d_g4_detsim.root

Cheers.

#16 Updated by David Adams about 4 years ago

Can you point me to the gen, g4, detsim and reco fcl files used for the sample? Thanks.

#17 Updated by David Adams about 4 years ago

For FD1x2x6 with single muons (not yet what Karl requested), I do see that dataprep affects the total memory used.

Full reco:

Events with large Vsize (Mbytes)        Vsize      Δ Vsize      RSS         Δ RSS
====================================================================================
  run: 20000014 subRun: 0 event: 8    1649.172     240.316     825.141     237.184
  run: 20000014 subRun: 0 event: 6    1569.898     250.289     745.324     246.637
  run: 20000014 subRun: 0 event: 4    1521.809     230.266     700.777     230.219
  run: 20000014 subRun: 0 event: 3    1471.781     180.238     647.152     176.617

Dataprep only:
Events with large Vsize (Mbytes)        Vsize      Δ Vsize      RSS         Δ RSS
====================================================================================
  run: 20000014 subRun: 0 event: 4    1459.555     190.254     630.758     187.145
  run: 20000014 subRun: 0 event: 8    1449.551     180.250     621.941     178.328
  run: 20000014 subRun: 0 event: 7    1449.531     180.230     621.543     177.930
  run: 20000014 subRun: 0 event: 3    1414.477     145.176     585.656     142.043

All but dataprep:
Events with large Vsize (Mbytes)         Vsize      Δ Vsize      RSS         Δ RSS
=====================================================================================
  run: 20000014 subRun: 0 event: 7     1347.062     -44.836     525.684     -43.113
  run: 20000014 subRun: 0 event: 10    1306.699     27.953      485.375     27.984
  run: 20000014 subRun: 0 event: 8     1279.906     -67.156     458.551     -67.133
  run: 20000014 subRun: 0 event: 9     1278.746     -0.957      457.391     -0.957

#18 Updated by David Adams about 4 years ago

It would be helpful to have a memory report for every event rather than those with largest memory and Delta-memory (which is anyway incorrect--see #13873). To do this, add "--memcheck-db mem.dat" to the lar command line and then read the mysql file with

sqlite3 -header -column mem.dat 'select * from EventInfo;'

Thanks to Kyle Knoepfel for this information.

#19 Updated by David Adams about 4 years ago

Here are some all-event memory reports, again for FD 1x2x6.

For full reco:

Run         Subrun      Event       Vsize        DeltaVsize  RSS           DeltaRSS    
----------  ----------  ----------  -----------  ----------  ------------  ------------
20000014    0           1           1426.453125  344.453125  597.64453125  339.47265625
20000014    0           2           1446.71875   155.253906  621.04296875  152.83984375
20000014    0           3           1471.757812  180.238281  644.421875    176.03515625
20000014    0           4           1521.777343  230.257812  698.2265625   229.8203125 
20000014    0           5           1291.519531  0           468.47265625  0           
20000014    0           6           1566.824218  275.304687  742.4609375   273.98828125
20000014    0           7           1665.492187  260.359375  838.765625    256.66015625
20000014    0           8           1645.46875   240.335937  819.62109375  237.5078125 
20000014    0           9           1405.132812  0           582.13671875  0.00390625  
20000014    0           10          1405.132812  0           582.1640625   0.02734375

For DataPrep only:

Run         Subrun      Event       Vsize        DeltaVsize  RSS          DeltaRSS  
----------  ----------  ----------  -----------  ----------  -----------  ----------
20000014    0           1           1401.609375  341.96875   573.2109375  335.375   
20000014    0           2           1409.292968  140.175781  582.125      138.769531
20000014    0           3           1414.308593  145.191406  585.2148437  141.859375
20000014    0           4           1459.371093  190.253906  629.7421875  186.386718
20000014    0           5           1269.117187  0           443.3554687  0         
20000014    0           6           1529.421875  260.304687  701.703125   258.347656
20000014    0           7           1444.320312  175.203125  618.0390625  174.683593
20000014    0           8           1449.335937  180.21875   619.3554687  176       
20000014    0           9           1269.117187  0           443.3554687  0         
20000014    0           10          1404.304687  135.1875    576.3007812  132.945312

For all but DataPrep:

Run         Subrun      Event       Vsize         DeltaVsize  RSS          DeltaRSS   
----------  ----------  ----------  ------------  ----------  -----------  -----------
20000014    0           1           1166.8671875  14.1640625  342.0234375  21.98046875
20000014    0           2           1198.0898437  21.4882812  374.5625     25.66796875
20000014    0           3           1218.2773437  20.1875     394.7617187  20.19921875
20000014    0           4           1218.265625   -0.0117187  394.8164062  0.0546875  
20000014    0           5           1218.265625   0           394.8164062  0          
20000014    0           6           1387.8671875  169.601562  563.78125    168.9648437
20000014    0           7           1347.6445312  -44.417968  524.2226562  -42.6640625
20000014    0           8           1280.6601562  -66.859375  457.2578125  -66.8398437
20000014    0           9           1280.578125   0           457.1796875  0.00390625 
20000014    0           10          1307.4296875  26.8515625  484.0585937  26.87890625

And doing nothing (no producers):

Run         Subrun      Event       Vsize         DeltaVsize  RSS           DeltaRSS  
----------  ----------  ----------  ------------  ----------  ------------  ----------
20000014    0           1           1055.3671875  0           235.51171875  0.0234375 
20000014    0           2           1126.6054687  0           300.765625    0         
20000014    0           3           1126.6054687  0           300.765625    0         
20000014    0           4           1126.6054687  0           302.4296875   0         
20000014    0           5           1126.6054687  0           302.4296875   0         
20000014    0           6           1131.609375   0           304.16796875  0         
20000014    0           7           1129.4179687  0           304.16796875  0         
20000014    0           8           1129.3867187  0           304.16796875  0         
20000014    0           9           1129.3867187  0           304.16796875  0         
20000014    0           10          1129.3867187  0           304.16796875  0

Kyle tells me the Vsize number is after event processing: "The Vpeak and RSS values correspond to the current memory usage, captured and recorded at the after-event boundary.".
Presumably Vsize-DeltaVsize is the memory usage before the event is processed.

#20 Updated by David Adams about 4 years ago

I did find a small memory leak in DataPrep. In the case now ROIs were found in a channel I was creating a Wire but not putting it in the event store where it would be deleted at the end of event processing. This will soon go into dunetpc develop.

The above memory reports (taken after fixing that) show no evidence of a large memory leak. It should be possible to reduce the memory usage by processing one APA or one APA plane at a time. I will have to check if the geometry supports this yet.

#21 Updated by David Adams about 4 years ago

Karl:

The above fix is in the current version (4f57b804e75ff0daef0839ffbde26e231965b001) of dunetpc. Can you check if this resolves your problem?

#22 Updated by David Adams about 4 years ago

Karl and Tingjun:

Can you tell me which for which release the memory consumption was significantly less? I would like to generate and compare with the above values.

Is or was someone monitoring this for every release?

#23 Updated by David Adams about 4 years ago

Issue #13877 proposes a way to reduce dataprep memory consumption for detectors with many APAs. It may provide the resolution for this issue.

#24 Updated by Tingjun Yang almost 4 years ago

I observe large memory usage when I try to run the reconstruction fcl file on data and save intermediate product caldata:noiseRemoved. It uses more than 8 G memory so all grid jobs failed. Here is the fcl file I used:
/dune/data/users/tjyang/lbne35t/standard_reco_dune35tdata.fcl

I will try to see if there is a memory leak using valgrind.

#25 Updated by Tingjun Yang almost 4 years ago

Additional information.

When setting services.RawDigitPrepService.DoIntermediateStates: false
MemoryTracker General SUMMARY (all numbers in units of Mbytes)

Peak virtual memory usage (VmPeak)  : 1683.73 Mbytes
  Peak resident set size usage (VmHWM): 805.676 Mbytes

Modules with large Vsize (Mbytes) Vsize Δ Vsize RSS Δ RSS =============================================================================================
---------------------------------------------------------------------------------------------
reco:caldata:DataPrepModule
run: 16256 subRun: 1 event: 1 1543.211 428.730 663.070 363.078
run: 16256 subRun: 1 event: 7 1524.309 0.109 712.145 0.109
run: 16256 subRun: 1 event: 5 1524.199 0.656 712.027 0.656
run: 16256 subRun: 1 event: 6 1524.199 0 712.035 0
---------------------------------------------------------------------------------------------

When setting services.RawDigitPrepService.DoIntermediateStates: true
physics.producers.caldata.IntermediateStates: ["noiseRemoved"]

MemoryTracker General SUMMARY (all numbers in units of Mbytes)

Peak virtual memory usage (VmPeak)  : 2681.46 Mbytes
  Peak resident set size usage (VmHWM): 1809.75 Mbytes

Modules with large Vsize (Mbytes) Vsize Δ Vsize RSS Δ RSS =============================================================================================
reco:caldata:DataPrepModule
run: 16256 subRun: 1 event: 9 2599.793 231.289 1721.211 165.656
run: 16256 subRun: 1 event: 8 2518.285 229.879 1639.641 164.242
run: 16256 subRun: 1 event: 7 2437.480 232.137 1558.723 166.352
run: 16256 subRun: 1 event: 6 2356.418 241.109 1477.660 175.152
---------------------------------------------------------------------------------------------

There does seem to be a memory leak in DataPrep when we save intermediate product.

Tingjun

Tingjun Yang wrote:

I observe large memory usage when I try to run the reconstruction fcl file on data and save intermediate product caldata:noiseRemoved. It uses more than 8 G memory so all grid jobs failed. Here is the fcl file I used:
/dune/data/users/tjyang/lbne35t/standard_reco_dune35tdata.fcl

I will try to see if there is a memory leak using valgrind.

#26 Updated by David Adams almost 4 years ago

  • Status changed from Assigned to Feedback
  • Assignee changed from Thomas Warburton to Tingjun Yang

Tingjun:

As discussed above, the log report of "large" memory usage is buggy. Could you follow the instructions for comment 18 above (#13749#note-18) and provide the report for all (and at least 20) events? Thanks.

#27 Updated by Thomas Warburton almost 4 years ago

Hi David,

Apologies for not replying for such a long time, I got overwhelmed with moving...At the end of last week and this morning I tried re-running dataprep on the 35 ton CRY and Far Detector geometries.

The good news is that I now definitely agree that the 35 ton geometry is fine. I am going to push through the last of the milliblock MCC samples now...
However for the Far Detector I am still seeing a problem even after the fix. I have been using v06_07_00 so that should have your fix applied in it.

I ran standard_detsim_dune10kt.fcl on this file /pnfs/dune/scratch/dunepro/v06_04_01/g4/prodMUSUN_dune10kt/10109056_90/prodMUSUN_dune10kt_52_20160826T220138_gen_d78e3ff0-bbcf-4208-b9bd-ffe46b21a2dc_g4.root to make sure that I had the latest version of detsim just in case that was somehow affecting things.

I then ran the following fcl files on the output of that file ( first twenty events ).
standard_reco_dune10kt.fcl

Run         SubRun      Event       Vsize       DeltaVsize    RSS            DeltaRSS   
----------  ----------  ----------  ----------  ------------  -------------  -----------
20000031    53          5201        2730.625    1594.7265625  1962.18359375  1591.984375
20000031    53          5202        4427.80078  1697.171875   3661.96875     1699.261718
20000031    53          5203        2724.26562  0             1958.62890625  0.0078125  
20000031    53          5204        2724.26562  0             1958.6328125   0.00390625 
20000031    53          5205        2724.26562  0             1958.63671875  0.00390625 
20000031    53          5206        2724.26562  0             1958.640625    0.00390625 
20000031    53          5207        2724.26562  0             1958.6484375   0.0078125  
20000031    53          5208        4426.26562  1702          3660.51953125  1701.871093
20000031    53          5209        2724.28906  0             1958.734375    0.00390625 
20000031    53          5210        2724.28906  0             1958.73828125  0.00390625 
20000031    53          5211        2724.28906  0             1958.74609375  0.0078125  
20000031    53          5212        2724.28906  0             1958.75        0.00390625 
20000031    53          5213        2724.28906  0             1958.75390625  0.00390625 
20000031    53          5214        2724.28906  0             1958.7578125   0.00390625 
20000031    53          5215        4460.93359  1736.6445312  3695.25390625  1736.496093
20000031    53          5216        4468.16406  1743.875      3702.48828125  1743.722656
20000031    53          5217        2724.28906  0             1958.7890625   0.0078125  
20000031    53          5218        4499.22656  1774.9375     3734.48828125  1775.699218
20000031    53          5219        4469.86718  1745.578125   3706.734375    1747.050781
20000031    53          5220        4330.15625  1605.8671875  3567.125       1605.800781

standard_reco_dune10kt.fcl ( Just data prep )
Run         SubRun      Event       Vsize         DeltaVsize     RSS           DeltaRSS   
----------  ----------  ----------  ------------  -------------  ------------  -----------
20000031    53          5201        2544.6328125  1594.73828125  1935.2109375  1591.640625
20000031    53          5202        4235.5234375  1690.88671875  3626.3125     1690.785156
20000031    53          5203        2538.0507812  0              1929.625      0.59375    
20000031    53          5204        2538.0507812  0              1930.4453125  0.8203125  
20000031    53          5205        2545.8398437  7.7890625      1938.9140625  8.46484375 
20000031    53          5206        2538.0507812  -7.7890625     1931.1289062  -7.78515625
20000031    53          5207        2538.0507812  0              1931.2421875  0.11328125 
20000031    53          5208        4228.3320312  1690.28125     3621.3359375  1690.09375 
20000031    53          5209        2538.0507812  0              1931.2421875  0          
20000031    53          5210        2538.0507812  0              1931.2421875  0          
20000031    53          5211        2538.0507812  0              1931.2421875  0          
20000031    53          5212        2538.0507812  0              1931.2421875  0          
20000031    53          5213        2543.8242187  5.7734375      1937.015625   5.7734375  
20000031    53          5214        2538.0507812  -5.7734375     1931.2421875  -5.7734375 
20000031    53          5215        4260.2929687  1722.2421875   3653.328125   1722.085937
20000031    53          5216        4264.3007812  1726.25        3657.3203125  1726.078125
20000031    53          5217        2538.0507812  0              1931.2421875  0          
20000031    53          5218        4295.0898437  1757.0390625   3688.1445312  1756.902343
20000031    53          5219        4263.6171875  1725.56640625  3656.6445312  1725.402343
20000031    53          5220        4123.515625   1585.46484375  3516.6328125  1585.390625

standard_reco_dune10kt.fcl ( All but data prep )
Run         SubRun      Event       Vsize        DeltaVsize  RSS           DeltaRSS  
----------  ----------  ----------  -----------  ----------  ------------  ----------
20000031    53          5201        1199.015625  0           435.05859375  0.91796875
20000031    53          5202        1248.445312  37.5585937  488.03125     40.7851562
20000031    53          5203        1247.875     0           487.66796875  0.00390625
20000031    53          5204        1247.875     0           487.671875    0.00390625
20000031    53          5205        1247.875     0           487.6796875   0.0078125 
20000031    53          5206        1247.875     0           487.68359375  0.00390625
20000031    53          5207        1247.875     0           487.69140625  0.00390625
20000031    53          5208        1264.308593  16.4335937  506.640625    18.9492187
20000031    53          5209        1277.4375    0           519.93359375  0         
20000031    53          5210        1277.4375    0           519.93359375  0         
20000031    53          5211        1277.4375    0           519.93359375  0         
20000031    53          5212        1277.4375    0           519.93359375  0         
20000031    53          5213        1277.4375    0           519.93359375  0         
20000031    53          5214        1277.4375    0           519.93359375  0         
20000031    53          5215        1271.476562  -5.9609375  514.546875    -5.3867187
20000031    53          5216        1271.476562  -1.484375   514.5625      -1.3554687
20000031    53          5217        1271.714843  0           514.80078125  0         
20000031    53          5218        1273.664062  1.94921875  516.15625     1.35546875
20000031    53          5219        1275.976562  0           518.625       0         
20000031    53          5220        1269.640625  -6.3359375  512.7265625   -5.8984375

Whilst doing this I noted that the showering modules were using a lot of memory so I removed them to isolate any increase which data prep may be causing.

As can be seen from this data prep is affecting the memory usage still, perhaps this brings into focus the need for Issue 13877 to be resolved. Looking through the event display, the events which have large memory usage for data prep are the ones where hits are stored.

The last time which I pushed reconstruction all the way through was for the last MCC challenge which used LArSoft version v05_11_00.

Cheers,
Karl

#28 Updated by David Adams almost 4 years ago

I have just pushed to dunetpc develop a modification so that data prep now processes each APA separately. This should significantly reduce memory usage in the data prep step where there are many APAs.

#29 Updated by Thomas Warburton almost 4 years ago

  • Status changed from Feedback to Resolved

Hi David,

I just tried running the reconstruction over a newly generated sample of MUSUN files in the far detector (using version v06_15_01). I forgot to run the memset command, but the simple memory usage was much much better - below. This was done for the whole reconstruction chain ( taking out pandora, showering and some of the later things for PMTrack as I got some startup errors for them ), and as you can see the memory usage is well below 2 GB.

Events increasing Vsize (Mbytes) Vsize Δ Vsize RSS Δ RSS =====================================================================================
run: 20000031 subRun: 0 event: 1 1420.867 249.547 622.152 247.395
run: 20000031 subRun: 0 event: 12 1431.523 7.750 638.926 7.746
run: 20000031 subRun: 0 event: 7 1430.289 6.516 635.914 7.258
run: 20000031 subRun: 0 event: 4 1423.840 2.449 627.309 2.516
run: 20000031 subRun: 0 event: 2 1421.492 0.621 624.871 2.090

Events with large Vsize (Mbytes) Vsize Δ Vsize RSS Δ RSS =====================================================================================
run: 20000031 subRun: 0 event: 12 1431.523 7.750 638.926 7.746
run: 20000031 subRun: 0 event: 13 1430.949 -0.574 638.352 -0.574
run: 20000031 subRun: 0 event: 7 1430.289 6.516 635.914 7.258
run: 20000031 subRun: 0 event: 4 1423.840 2.449 627.309 2.516
run: 20000031 subRun: 0 event: 5 1423.773 -0.066 628.031 0.723

====================================================================================================

Thankyou!

I think that this means for now at least we can consider the ticket resolved.

#30 Updated by David Adams almost 4 years ago

  • Status changed from Resolved to Closed

Thanks. I close the ticket

Also available in: Atom PDF