PTP/meetings/November 2006

November 2006 Meeting

Date: November 14, 2006 Time: 1:00 PM EST

Call in information:

Toll Free: 1 866 807-2880
Local: (505) 606-1028

Agenda

1. PTP Status
   • 1.1 release update
2. Project Reports
   • PLDT update
   • Performance tools update
   • Fortran update
   • Resource manager/runtime support update
   • Remote services update
3. Tutorial Report
4. Other business
5. Next meeting

Attendees

1. Greg Watson, LANL
2. Craig Rasmussen, LANL
3. Beth Tibbitts, IBM
4. Evelyn Duesterwald, IBM
5. Yuan Zhang, IBM
6. Don Pazel, Google
7. Diglio Simoni, RTI
8. Wyatt Spear, U. Oregon
9. Alan Morris, U. Oregon
10. Tianchao Li, Technische Universität München (by phone)
11. Doug Schaefer, QNX (by phone)
12. Chris Recoskie, IBM (by phone)

Meeting Notes

Meeting was held at Supercomputing2006 convention site, with the largest ever gathering (in person) of PTP participants! Brief introduction of all attendees.

• Greg: Tutorial given at SC06 on Sunday. Over 30 attendees.

• PTP 1.1 status - Greg:
  • Release 1.1 is close (we're at "RC1" effectively). Release of 1.1 by end of year. See builds page; will do change logs.
  • Changes include:
    1. bug fixes to parallel debugger
    2. support for OMPI 1.2
    3. OpenMP support in PLDT
  • TODO items for release: release notes, pre-req versions information, etc.

• Resource Manager Status - Greg: Craig, Randy and Greg working on this. Integrating resource managment and remote support now.
  • Now supports "null resource manager" which is OpenMPI. Next will support a real res.mgr. e.g. LoadLeveler

• Remote support - Craig: working on design and integration with resource manager. Will create requirement doc and then work on design docs.
  • RSE 1.0 is out; PTP doing similar things to RSE but other function needed that is not available in RSE. Need to talk to David Dykstal and Martin Oberhuber.

• PTP for teaching - Evelyn reported interest in using PTP to teach HPC course at Calvin College in Michigan - Joel Adams. He has OpenMPI installed now, in preparation. Could take info to teaching conference to share new course curriculum. Greg said Univ. Toronto is setting up an HPC center to set up HPC courses. Greg's goal is to get the debugger into universities, to facilitate research in parallel debugging.

• Diglio - MOAB - meeting with David Jackson, Cluster Resources.

Should we form a PTP subgroup on resource management?

• IBM report - Beth reported IBM group working on LoadLeveler and IBM PE/POE (Parallel Environment) implementation for PTP.
  • PLDT work includes MPI Barrier analysis work, demo'd here at SC06 in the IBM booth.
  • PE (Performance Explorer visualization of trace data) is also being demo'd at the IBM booth here at SC06; hopefully to be contributed by PTP 2.0 (summer '07). Early access probably in Spring. PE uses XML file to describe visualization. Wyatt is interested in hooking up TAU traces to PE.

• TAU performance analysis framework for code instrumentation - Wyatt
  • Enhanced plugin setup for TAU; analysis tab to launch configuration - done by Alan Morris. Output stored in (Derby) database; Paraprof can be launched from Eclipse for visualization; can link back to source code in Eclipse. TAU cleans up instrumentation when done. Evelyn asked whether this conflicts with compiler instrumentation? Greg asked how much interaction in the plugin is TAU specific. Wyatt said most of it. But not hard to decouple. Alan: no other system has compiler wrapper scripts. Greg: eclipse how-to article on configuration of e.g. launch? Wyatt: will do some code cleanup.
  • Beth asked if anyone knew if Shirley Moore had made progress on the Eclipse plugin for performance instrumentation for ?PAPI? - no info. Wyatt needs feedback from CDT on when execution completes. Wyatt is demoing TAU at the LANL booth here at SC06.

• PLDT - MPI barrier analysis - Yuan - Barrier analysis to find syncronization errors. Quite precise, few spurious errors.
  • This is analysis specific to barriers, unlike the existing OpenMP concurrency analysis in PLDT. IBM plans to contribute generic analysis code to CDT, and MPI specific code to PTP. Yuan hopes to continue work on PLDT barrier analysis when she returns to U.Delaware. There is no specific UI to the call graph and concurrency graph that was developed for the MPI barrier analysis - only used in analysis, not for display to user. The MPI barrier analysis contains its own UI for displaying results to the user.
  • Yuan/Beth asked for suggestions for display of directed graphs. Dan Quinlan with Rose Call graph / directed graph was suggested (not Eclipse however); Alan said he has used Java2D JGraph for analysis.

• Don is interested in participating in PTP again. Wonders if the amount of time will be enough. Interests include debugging, global state/condition, global snapshot, scalability issues; focues on dynamic backends. Global State: send/broadcast states. Take 'for' condition across the application, zoom in when it happens.

• Fortran - Craig: Fortran 2003 ANTLR grammar complete. Chris Rickett at LANL has taken over this portion of the project. About 40 unit tests have been done. AST generation is next (next week). Dan Quinlan / Rose project / Static analaysis for C++ and looking at Fortran. Hope to generate AST.
  • Chris Recoskie mentioned that ANTLR might not be usable by IBM - IP status is unclear. Looking at other parser generators in case can't use ANTLR. Chris will keep us informed.
  • Doug said he could take a beta release thru the release process to see - will do it for C# and let us know. He has contacts in Eclipse legal department in Ottawa. Alternatives: LPG LALR parser?

• PTP move from the Technology Project to the Tools project of Eclipse - Doug discussed the proposed move. Proposed since PTP is more mature and beyond the incubation stage now. Need to schedule a move review - similar to a release review.
  • Greg wants to talk to Doug and the Tools PMC re: managing release process.

Meeting adjourned: 9 PTP'ers in person!

Next meeting: December 12 (2nd Tuesday in December)