Shared Memory Parallelism.ppt

Origin Optimisation and Parallelisation Training March 2000 Chapter 11. SMP Parallelisation Origin Optimisation and Parallelisation Training Shared Memory Parallelism Programming in the SGI Environment Parallel Programming on SGI Automatic parallelization with APC Fortran and C compiler option: -apo HPF: pghpf (Portland Group), xHPF (third party) parallelization by inserting OpenMP compiler directives (using the -mp compiler option) Fortran !$omp parallel, !$omp do C #pragma omp parallel Low-Level parallelism Posix standard Pthreads SGI library sproc() UNIX native fork() Sequent library m_fork() Message Passing PVM, MPI shmem Definition of number of parallel threads (man pe_environ) SMP (APC and OpenMP, sproc): setenv OMP_NUM_THREADS n MPI: mpirun -np n prog SMP Programming The parallelism in the code: can be found by the compiler can be asserted by the programmer with directives to help the compiler to analyze the code compiler analyzes data access rules based on the additional information presented can be forced by the programmer with directives/pragmas compiler has to be told about data access rules (shared/private) explicit synchronization is often necessary to access the shared data can be forced by the programmer with library calls POSIX Pthreads Automatic Parallelization Automatic Parallelization Automatically Parallelizable Code linear assignments: reductions: triangular assignments NOT parallelizable: early exits Automatic Parallelization Automatic Parallelization Automatic Parallelization Compiler Assertions Assertions help the compiler to deal with data dependencies that cannot be analyzed based on program semantics alone: C*$* (NO) CONCURRENTIZE #pragma (no) concurrentize (not) to apply parallel optimizations, the scope is the entire compilation unit C*$* ASSERT DO (SERIAL) #pragma serial C*$* ASSERT DO (CONCURRENT) #pragma concurrent implement the required action for the loop (nest) immediately following the directive C*$* ASSER