MantisBT - Rosetta
View Issue Details
IDProjectCategoryView StatusDate SubmittedLast Update
0000031Rosetta[All Projects] Crashpublic2012-06-27 22:022012-08-03 11:49
Reporterdelucasl 
Assigned Todelucasl 
PrioritylowSeveritycrashReproducibilityrandom
StatusassignedResolutionopen 
PlatformAll platformsOSAnyOS VersionAny
Product VersionTrunk 
Fixed in Version 
Application(s) Affectedanything supporting jd2
Command Line Usedwhatever.mysqlmpi.linuxgccrelease @flags.txt -jd2:mpi_fast_nonblocking_output
Developer Options
Fixed in SVN Version
Summary0000031: Using DatabaseFilters and mysql/postgre and non-blocking mpi at the same time causes a race condition
DescriptionDatabaseFilters write a structure and then if necessary delete the structure being filtered out. In normal operation this is fine, as the MPIWorkPoolJobDistributor blocks so only one node can output at a time. Blocking is not strictly necessary when outputting to a database server, and can cause bottlenecking in short protocols. However, if blocking is disabled using the -jd2:mpi_fast_nonblocking_output, two processes can try to simultaniously delete the same pose, which causes a crash when the delete option fails.
Steps To Reproducerun any jd2 supported application outputting to a database server with a large number of MPI managed nodes, a large nstruct and jd2:mpi_fast_nonblocking_output turned on. It will crash at some point when the race condition occurs.
Additional Information
TagsNo tags attached.
Relationships
Attached Files
Issue History
Date ModifiedUsernameFieldChange
2012-06-27 22:02delucaslNew Issue
2012-06-27 22:02delucaslStatusnew => assigned
2012-06-27 22:02delucaslAssigned To => delucasl
2012-08-03 11:49delucaslNote Added: 0000046
2012-08-03 11:49delucaslPrioritynormal => low

Notes
(0000046)
delucasl   
2012-08-03 11:49   
I'm lowering this to low priority because it's actually quite tricky to fix and it's not really that big a deal to just use the nonblocking output for the time being.