IV01042: PE 5.2 HANGS WITH MPI PERSISTENT COMMUNICATION

Readme and Release notes for release 5.2.2.4 PE-5.2.2.4-power-AIX-61
Readme and Release notes for release 5.2.2.4 PE-5.2.2.4-power-Linux
Readme and Release notes for release 5.2.2.4 PE-5.2.2.4-x86-Linux
Readme and Release notes for release 5.2.2.4 PE-5.2.2.4-power-AIX-53
Readme and Release notes for release 5.2.2.5 PE-5.2.2.5-power-AIX-53
Readme and Release notes for release 5.2.2.5 PE-5.2.2.5-power-Linux
Readme and Release notes for release 5.2.2.5 PE-5.2.2.5-x86-Linux
Readme and Release notes for release 5.2.2.5 PE-5.2.2.5-power-AIX-61
Readme and Release notes for release 5.2.2.6 PE-5.2.2.6-power-AIX-53
Readme and Release notes for release 5.2.2.6 PE-5.2.2.6-power-AIX-61
Readme and Release notes for release 5.2.2.6 PE-5.2.2.6-power-Linux
Readme and Release notes for release 5.2.2.6 PE-5.2.2.6-x86-Linux
Readme and Release notes for release 5.2.2.7 PE-5.2.2.7-power-AIX-53
Readme and Release notes for release 5.2.2.7 PE-5.2.2.7-power-AIX-61
Readme and Release notes for release 5.2.2.7 PE-5.2.2.7-power-Linux
Readme and Release notes for release 5.2.2.7 PE-5.2.2.7-x86-Linux
Readme and Release notes for release 5.2.2.8 PE-5.2.2.8-x86-Linux
Readme and Release notes for release 5.2.2.8 PE-5.2.2.8-power-Linux

APAR status

• Closed as program error.

Error description

• The customer upgraded to PE 5.2 and now their
  applications hangs. The problem is that the MPI
  persistent communication has a problem in PE 5.2
  in an internal function which computes the rank to
  task id and does not handle  MPI_PROC_NULL correctly.
  Somehow it gets set to 0, so we have other task
  sitting on a persistent call waiting for task 0
  but task 0 is sitting on a MPI_Barrier call
  where all tasks should participate.
  

Local fix

Problem summary

• The application calls persistent functions with
  MPI_PROC_NULL. When we upgrade from 5.1.1 to 5.2.1, MPI
  internal function that computes the rank to task id does not
  handle MPI_PROC_NULL correctly, somehow it gets set to 0, so
  we have other task on a persistent call waiting for task 0
  but task 0 is on a MPI_Barrier call where all tasks should
  participate.
  

Problem conclusion

• When MPI_PROC_NULL is passed to the computation function of
  rank to task id, it should be set to MPCI_PROC_NULL, then
  the communication layer will handle correctly.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

Fix information

• Fixed component name

  PARALLEL ENVIR.

• Fixed component ID

  5765PEA00

Applicable component levels

• R522 PSY U859445

     UP13/05/02 I 1000

PTF to Fileset Mapping

• U844522 ppe.poe 5.2.2.4

• U844522 ppe.poe 5.2.2.4

• U846782 ppe.poe 5.2.2.5

• U846782 ppe.poe 5.2.2.5

• U849082 ppe.poe 5.2.2.6

• U849082 ppe.poe 5.2.2.6

• U851494 ppe.poe 5.2.2.7

• U851494 ppe.poe 5.2.2.7

• U859445 ppe.poe 5.2.2.8

• U859445 ppe.poe 5.2.2.8