MantisBT - ParaView
View Issue Details
IDProjectCategoryView StatusDate SubmittedLast Update
0006035ParaView(No Category)public2007-11-09 14:282009-05-13 13:59
ReporterDavid Karelitz 
Assigned ToBerk Geveci 
PriorityhighSeveritycrashReproducibilityalways
StatusclosedResolutionduplicate 
PlatformOSOS Version
Product Version 
Target Version3.4Fixed in Version3.4 
Project
Topic Name
Type
Summary0006035: (MARS-immediate) 0th process using too much memory
DescriptionWhen opening and performing operations on a large CTH AMR dataset, the 0th node uses way too much memory. Order 3-4 GB on the 0th node vs 640 MB on other nodes. There was a CTH reader, CellToPoint, Contour, Calculator, TimeToText, and TextSource in the pipeline and we tried to save a movie and disconnect.
Steps To Reproduce
Additional Information
TagsNo tags attached.
Relationships
duplicate of 0006773closed Utkarsh Ayachit Use Broadcast to propagate commands from root sever node to other server nodes 
Attached Files
Issue History
Date ModifiedUsernameFieldChange
2007-11-09 14:28David KarelitzNew Issue
2007-11-15 16:45David KarelitzNote Added: 0009690
2007-11-15 16:50David KarelitzNote Added: 0009691
2007-11-19 18:19David KarelitzSummary0th process using too much memory => (MARS-immediate) 0th process using too much memory
2007-12-05 13:13Alan ScottStatusbacklog => tabled
2007-12-05 13:13Alan ScottAssigned To => Berk Geveci
2008-01-10 11:11David KarelitzNote Added: 0010104
2008-03-07 07:56Berk GeveciTarget Version => MARS
2008-03-07 08:03Berk GeveciCategory => 3.4
2008-04-08 13:01Ken MorelandRelationship addedduplicate of 0006773
2008-04-08 13:02Ken MorelandDuplicate ID0 => 6773
2008-04-08 13:02Ken MorelandStatustabled => @80@
2008-04-08 13:02Ken MorelandResolutionopen => duplicate
2008-04-08 13:02Ken MorelandNote Added: 0011159
2008-04-08 17:56Alan ScottStatus@80@ => closed
2008-04-08 17:56Alan ScottNote Added: 0011188
2009-05-13 13:58Utkarsh AyachitTarget VersionMARS => 3.4
2009-05-13 13:59Utkarsh AyachitFixed in Version => 3.4
2011-06-16 13:10Zack GalbreathCategory => (No Category)

Notes
(0009690)
David Karelitz   
2007-11-15 16:45   
There appear to be 2 issues with respect to memory usage. These are from a separate run than that in the description

1. When I clip by scalar value with an incorrect value (that results in more elements than the correct value), the memory is not released for the first incorrect clip value.

2. When I choose save and disconnect, memory usage on the 0th node jumps from 33% to 85% and uses an additional 2-4% per frame. This is on an 8 GB machine.

The dataset is AMRCTH running on 120 nodes
(0009691)
David Karelitz   
2007-11-15 16:50   
It's 4 GB of memory per node, not 8GB
(0010104)
David Karelitz   
2008-01-10 11:11   
When we switched the 0th node to a node with 12GB of memory, performance was an order of magnitude faster. The rest of the nodes were the same as before
(0011159)
Ken Moreland   
2008-04-08 13:02   
This problem is really caused by a bug in OpenMPI and the way that messages are passed from node 0 to the rest of the nodes. I opened a new bug that more specifically addresses the problem.
(0011188)
Alan Scott   
2008-04-08 17:56   
I agree with ken. This is an open mpi bug, not related to the bug description listed below. Solution is th change MPI_Send to MPI_Ssend, until the bug in the OpenMPI library is fixed. I believe the file is MPICommunication.cxx or something like that.