1. www.vacet.org E. WES BETHEL (LBNL), CHRIS JOHNSON (UTAH), KEN JOY (UC DAVIS), SEAN AHERN (ORNL), VALERIO PASCUCCI (LLNL), JONATHAN COHEN (LLNL), MARK DUCHAINEAU (LLNL), BERND HAMANN (UC DAVIS), CHARLES HANSEN (UTAH), DAN LANEY (LLNL), PETER LINDSTROM (LLNL), JEREMY MEREDITH (ORNL), GEORGE OSTROUCHOV (ORNL), STEVEN PARKER (UTAH), CLAUDIO SILVA (UTAH), XAVIER TRICOCHE (UTAH), ALLEN SANDERSON (UTAH), HANK CHILDS (LLNL) Lessons Learned From the MPI- Hybrid Parallelism for Streamlines on Large Multi-Core Clusters Project David Camp (IDAV)

2. www.vacet.org MPI-Hybrid Other VACET projects have shown good performance gains with MPI-Hybrid This project wanted to explore MPI-Hybrid style with two standard Streamlines algorithms, LOD and Static Domains Talk about some of the problems encountered and performance gains Other VACET projects have shown good performance gains with MPI-Hybrid This project wanted to explore MPI-Hybrid style with two standard Streamlines algorithms, LOD and Static Domains Talk about some of the problems encountered and performance gains

3. www.vacet.org Baseline test for MPI-Hybrid Original MPI test –Ran in 100 seconds, on 128 cores First MPI-Hybrid Test –Ran in ~20,000 seconds, on 128 cores Final MPI-Hybrid Test, After many fixes –Ran in 15 seconds, on 128 cores Original MPI test –Ran in 100 seconds, on 128 cores First MPI-Hybrid Test –Ran in ~20,000 seconds, on 128 cores Final MPI-Hybrid Test, After many fixes –Ran in 15 seconds, on 128 cores

4. www.vacet.org VTK At the heart of VTK pipeline is the data time stamp –This is used to drive their data flow model –Every action in VTK changes the data time stamp vtkTimeStamp::Modified() –A small test found Call ~1,000,000 times –Found a pthread_mutex_lock to protect the time stamp At the heart of VTK pipeline is the data time stamp –This is used to drive their data flow model –Every action in VTK changes the data time stamp vtkTimeStamp::Modified() –A small test found Call ~1,000,000 times –Found a pthread_mutex_lock to protect the time stamp

5. www.vacet.org Crashing in VTK VTK – Thread Safe? –Documentation said Thread Safe –Look like memory corruption VTK – Documents Say Thread Safe –But many function where defined “Not Thread Safe” –Some “This Method is Thread Safe if first called from a Single Thread and the dataset is not Modified” Real Answer is VTK is not Thread Safe –vtkObjectBase did not protect it reference count variable, so Data Concurrency was lost. –Memory was being delete before it life time had truly ended VTK – Thread Safe? –Documentation said Thread Safe –Look like memory corruption VTK – Documents Say Thread Safe –But many function where defined “Not Thread Safe” –Some “This Method is Thread Safe if first called from a Single Thread and the dataset is not Modified” Real Answer is VTK is not Thread Safe –vtkObjectBase did not protect it reference count variable, so Data Concurrency was lost. –Memory was being delete before it life time had truly ended

6. www.vacet.org C++ Exception Across Share Libraries Streamline code used an Exception to handle data boundary condition –Linux used a pthread_mutex_lock to handle this Execption Code was change to remove the exception Streamline code used an Exception to handle data boundary condition –Linux used a pthread_mutex_lock to handle this Execption Code was change to remove the exception

7. www.vacet.org VTK – Object Creation VTK forces you to use it’s New function –VTK uses a factory method pattern vtkObjectFactory –Used to override VTK classes with custom versions. –It used strcmp to match object –Strcmp was the most called function in the Streamlines test VTK forces you to use it’s New function –VTK uses a factory method pattern vtkObjectFactory –Used to override VTK classes with custom versions. –It used strcmp to match object –Strcmp was the most called function in the Streamlines test

8. www.vacet.org I/O Found that MPI I/O was better –They where doing multi-I/O operations by default by running four process per node Changed the Streamline code to thread I/O Found that MPI I/O was better –They where doing multi-I/O operations by default by running four process per node Changed the Streamline code to thread I/O

9. www.vacet.org Conclusion – Hard Work Pays Off Original MPI test –Run on Jaguar –100 seconds (10,000 Streamlines 128 cores) Original MPI test with code improvements –Run on Franklin –45 seconds (20,000 Streamlines 128 cores) MPI-Hybrid test –Run on Franklin –15 seconds (20,000 Streamlines 128 cores) Original MPI test –Run on Jaguar –100 seconds (10,000 Streamlines 128 cores) Original MPI test with code improvements –Run on Franklin –45 seconds (20,000 Streamlines 128 cores) MPI-Hybrid test –Run on Franklin –15 seconds (20,000 Streamlines 128 cores)

10. www.vacet.org Common Problems with Threads Concurrency –One or more threads access the same memory location without locking. Lock Contention –One thread blocks the progress of one or more other threads by holding a lock too long. VisIt’s Data Cache VisIt’s Timer Deadlock –When two or more threads wait for each other lock, indefinitely. Valgrind: DRD and helgrind –Tools to help find these types of problems. Concurrency –One or more threads access the same memory location without locking. Lock Contention –One thread blocks the progress of one or more other threads by holding a lock too long. VisIt’s Data Cache VisIt’s Timer Deadlock –When two or more threads wait for each other lock, indefinitely. Valgrind: DRD and helgrind –Tools to help find these types of problems.