MantisBT - ParaView
View Issue Details
IDProjectCategoryView StatusDate SubmittedLast Update
0007640ParaViewBugpublic2008-09-09 11:572009-09-11 14:10
ReporterUtkarsh Ayachit 
Assigned ToUtkarsh Ayachit 
PrioritynormalSeverityminorReproducibilityhave not tried
StatusclosedResolutionfixed 
PlatformOSOS Version
Product Version 
Target Version3.6.2Fixed in Version3.6.2 
Project
Topic Name
Type
Summary0007640: MPI Communication cleanup
DescriptionRunning ParaView as 16 processes cluster with openmpi (amber) is excruciatingly slow esp. in creating sources/filters/representations.
Steps To Reproduce
Additional InformationSuggested fixes:
* Don't use MPI_Ssend by default. Make it a CMake variable which can be set on clusters with buggy MPI_Send implementations.
* When broadcasting to all processess, the root node sequentially sends the message to all the satellite. Instead we should use a binary-tree like communication hierarchy to broadcast messages.
* Use buffering for SendStream(). The buffer will be flushed when GatherInformation or GetLastResult() is called. This will avoid several small messages being sent to the servers.
TagsNo tags attached.
Relationships
Attached Files
Issue History
Date ModifiedUsernameFieldChange
2008-09-09 11:57Utkarsh AyachitNew Issue
2009-02-17 13:27Utkarsh AyachitStatusbacklog => tabled
2009-02-17 13:27Utkarsh AyachitAssigned To => Utkarsh Ayachit
2009-02-17 13:27Utkarsh AyachitCategoryDevelopment => 3.6
2009-02-17 13:28Utkarsh AyachitNote Added: 0015039
2009-02-17 13:28Utkarsh AyachitAssigned ToUtkarsh Ayachit => Burlen
2009-04-22 13:23Utkarsh AyachitAssigned ToBurlen => Yumin Yuan
2009-05-13 13:45Utkarsh AyachitTarget Version => 3.6
2009-08-25 12:17Ken MorelandCategory => Bug
2009-08-25 12:17Ken MorelandTarget Version3.6 => 3.6.2
2009-08-31 08:56Utkarsh AyachitNote Added: 0017241
2009-08-31 09:36Utkarsh AyachitAssigned ToYumin Yuan => Utkarsh Ayachit
2009-08-31 09:41Utkarsh AyachitNote Added: 0017239
2009-08-31 09:43Utkarsh AyachitNote Added: 0017240
2009-08-31 09:43Utkarsh AyachitNote Edited: 0017239
2009-09-01 09:55Utkarsh AyachitNote Added: 0017273
2009-09-01 09:55Utkarsh AyachitStatustabled => @80@
2009-09-01 09:55Utkarsh AyachitFixed in Version => 3.6.2
2009-09-01 09:55Utkarsh AyachitResolutionopen => fixed
2009-09-11 14:10Alan ScottNote Added: 0017379
2009-09-11 14:10Alan ScottStatus@80@ => closed

Notes
(0015039)
Utkarsh Ayachit   
2009-02-17 13:28   
(2) and (3) are already fixed. We just need to add a Cmake option to enable ssend.
(0017241)
Utkarsh Ayachit   
2009-08-31 08:56   
/cvsroot/ParaView3/ParaView3/vtkPVConfig.h.in,v <-- vtkPVConfig.h.in
new revision: 1.9; previous revision: 1.8
(0017239)
Utkarsh Ayachit   
2009-08-31 09:41   
(edited on: 2009-08-31 09:43)
Adding new advanced cmake variable PARAVIEW_USE_MPI_SSEND
(off by default) to make ParaView use MPI_Ssend, instead of MPI_Send.

/cvsroot/ParaView3/ParaView3/CMakeLists.txt,v <-- CMakeLists.txt
new revision: 1.110; previous revision: 1.109
/cvsroot/ParaView3/ParaView3/Servers/Common/vtkProcessModule.cxx,v <-- Servers/Common/vtkProcessModule.cxx
new revision: 1.95; previous revision: 1.94


Pending commit to 3.6 branch.
(0017240)
Utkarsh Ayachit   
2009-08-31 09:43   
Reminder sent to: Alan Scott

Alan,

Make sure that you set the new PARAVIEW_USE_MPI_SSEND to ON before building your binaries on the machines with MPI_Send buffer overflow issues.
(0017273)
Utkarsh Ayachit   
2009-09-01 09:55   
/cvsroot/ParaView3/ParaView3/CMakeLists.txt,v <-- CMakeLists.txt
new revision: 1.97.2.7; previous revision: 1.97.2.6
/cvsroot/ParaView3/ParaView3/Servers/Common/vtkProcessModule.cxx,v <-- Servers/Common/vtkProcessModule.cxx
new revision: 1.92.2.2; previous revision: 1.92.2.1

/cvsroot/ParaView3/ParaView3/vtkPVConfig.h.in,v <-- vtkPVConfig.h.in
new revision: 1.5.2.1; previous revision: 1.5
(0017379)
Alan Scott   
2009-09-11 14:10   
Blackrose seems fine. I tested on Dave's large CTH dataset. If we get hangs on larger stuff, be sure to test this first. Tested remote server, 64 pvservers.