PMIx: Debuggers and Fabric Support
- 1. PMIx: Debuggers and Fabric
- 2. Debugger/Tool Features • Co-launch/co-location of daemons At initial app spawn • Co-launch Upon attach • Spawn w/co-location • Launch control Stop-on-exec, stop-in-init, stop-in-app Release method spec • Forwarding of IO To/from debugger daemons To/from app being debugged • Query support Global and local proctable PMIx server and application procs • Direct/indirect launch support Forward, set/unset/modify envars (e.g., LD_PRELOAD) Launcher directives • Modify local fork/exec agent • Replace launcher daemons
- 5. RM PMIx Server RM PMIx Server mpiexec PMIx Server Indirect Launch Launch Directives Tools: only connect to one server at a time
- 7. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Attach to Running Job Launch Daemons mpid
- 8. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Attach to Running Job mpid
- 9. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Attach to Running Job Launch Daemons mpid Direct or Indirect Launch
- 10. Web Links • https://pmix.org Home page for PMIx • https://pmix.org/pmix-standard/ Standards process, links to headers, RFCs Roles/responsibilities of RMs and other system management components • How-to guides https://pmix.org/support/how-to/example-direct-launch- debugger-tool/ https://pmix.org/support/how-to/example-indirect- launch-debugger-tool/
- 11. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn Assigning Procs->Daemons Query: • Local proctable • Local rank PMIX_DEBUG_JOB Assigned in launch data Query global proctable
- 12. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Assigning Procs->Daemons mpid Query: • Local proctable • Local rank PMIX_DEBUG_JOB Assigned in launch data
- 13. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Assigning Procs->Daemons mpid Query: • Local proctable • Local rank PMIX_DEBUG_JOB Assigned in launch data
- 14. General Capabilities • Specify mechanism for “hold” Query RM or launcher for support stop-on-exec, stop-in-init, or stop-in-app • Specify app release mechanism PMIx event, signal, … • Register for events Termination of debugger job and/or daemons Termination of app job and/or procs
- 15. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Query App Info mpid
- 16. RM PMIx Server RM PMIx Server Proc Proc stdout stderr Sent via PMIx Forwarding of Output Dbgr Dmn
- 17. RM PMIx Server RM PMIx Server Proc Proc stdin Sent via PMIx Forwarding Stdin PMIx Client Collects Tool Collects Dbgr Dmn
- 18. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Forwarding Stdin mpid stdin PMIx Client Collects Tool Collects Sent via PMIx
- 19. RM PMIx Server RM PMIx Server Proc Proc Dbgr Dmn PMIx Server mpiexec PMIx Server Forwarding Stdin mpid stdinPMIx Client Collects Tool Collects Sent via PMIx
- 20. Fabric Features • Harvest/forward envars Include/exclude • Get/set security keys • Get/forward setup “blob” • Deliver setup “blob” prior to local process spawn Support: • OmniPath (master) • TCP (branch)