1. ECE1747 Parallel Programming Shared Memory Multithreading Pthreads

2. Shared Memory proc1proc2 proc3procN Shared Memory Address Space All threads access the same shared memory data space.

3. Shared Memory (continued) Concretely, it means that a variable x, a pointer p, or an array a[] refer to the same object, no matter what processor the reference originates from. We have more or less implicitly assumed this to be the case in earlier examples.

4. Shared Memory proc1proc2proc3procN a

5. Distributed Memory - Message Passing The alternative model to shared memory. proc1proc2proc3procN mem1mem2mem3memN network aaaa

6. Shared Memory vs. Message Passing Same terminology is used in distinguishing hardware. For us: distinguish programming models, not hardware.

7. Programming vs. Hardware One can implement –a shared memory programming model –on shared or distributed memory hardware –(also in software or in hardware) One can implement –a message passing programming model –on shared or distributed memory hardware

8. Portability of programming models shared memory programming message passing programming distr. memory machine shared memory machine

9. Shared Memory Programming: Important Point to Remember No matter what the implementation, it conceptually looks like shared memory. There may be some (important) performance differences.

10. Multithreading User has explicit control over thread. Good: control can be used to performance benefit. Bad: user has to deal with it.

11. Pthreads POSIX standard shared-memory multithreading interface. Provides primitives for process management and synchronization.

12. What does the user have to do? Decide how to decompose the computation into parallel parts. Create (and destroy) processes to support that decomposition. Add synchronization to make sure dependences are covered.

13. General Thread Structure Typically, a thread is a concurrent execution of a function or a procedure. So, your program needs to be restructured such that parallel parts form separate procedures or functions.

14. Example of Thread Creation (contd.) main() pthread_ create(func) func()

15. Thread Joining Example void *func(void *) { ….. } pthread_t id; int X; pthread_create(&id, NULL, func, &X); ….. pthread_join(id, NULL); …..

16. Example of Thread Creation (contd.) main() pthread_ create(func) func() pthread_ join(id) pthread_ exit()

17. Sequential SOR for some number of timesteps/iterations { for (i=0; i<n; i++ ) for( j=1, j<n, j++ ) temp[i][j] = 0.25 * ( grid[i-1][j] + grid[i+1][j] grid[i][j-1] + grid[i][j+1] ); for( i=0; i<n; i++ ) for( j=1; j<n; j++ ) grid[i][j] = temp[i][j]; }

18. Parallel SOR First (i,j) loop nest can be parallelized. Second (i,j) loop nest can be parallelized. Must wait to start second loop nest until all processors have finished first. Must wait to start first loop nest of next iteration until all processors have second loop nest of previous iteration. Give n/p rows to each processor.

19. Pthreads SOR: Parallel parts (1) void* sor_1(void *s) { int slice = (int) s; int from = (slice*n)/p; int to = ((slice+1)*n)/p; for( i=from; i<to; i++) for( j=0; j<n; j++ ) temp[i][j] = 0.25*(grid[i-1][j] + grid[i+1][j] +grid[i][j-1] + grid[i][j+1]); }

20. Pthreads SOR: Parallel parts (2) void* sor_2(void *s) { int slice = (int) s; int from = (slice*n)/p; int to = ((slice+1)*n)/p; for( i=from; i<to; i++) for( j=0; j<n; j++ ) grid[i][j] = temp[i][j]; }

21. Pthreads SOR: main for some number of timesteps { for( i=0; i<p; i++ ) pthread_create(&thrd[i], NULL, sor_1, (void *)i); for( i=0; i<p; i++ ) pthread_join(thrd[i], NULL); for( i=0; i<p; i++ ) pthread_create(&thrd[i], NULL, sor_2, (void *)i); for( i=0; i<p; i++ ) pthread_join(thrd[i], NULL); }

22. Summary: Thread Management pthread_create(): creates a parallel thread executing a given function (and arguments), returns thread identifier. pthread_exit(): terminates thread. pthread_join(): waits for thread with particular thread identifier to terminate.

23. Summary: Program Structure Encapsulate parallel parts in functions. Use function arguments to parameterize what a particular thread does. Call pthread_create() with the function and arguments, save thread identifier returned. Call pthread_join() with that thread identifier.

24. Pthreads Synchronization Create/exit/join –provide some form of synchronization, –at a very coarse level, –requires thread creation/destruction. Need for finer-grain synchronization –mutex locks, –condition variables.

25. Use of Mutex Locks To implement critical sections. Pthreads provides only exclusive locks. Some other systems allow shared-read, exclusive-write locks.

26. Barrier Synchronization A wait at a barrier causes a thread to wait until all threads have performed a wait at the barrier. At that point, they all proceed.

27. Implementing Barriers in Pthreads Count the number of arrivals at the barrier. Wait if this is not the last arrival. Make everyone unblock if this is the last arrival. Since the arrival count is a shared variable, enclose the whole operation in a mutex lock-unlock.

28. Implementing Barriers in Pthreads void barrier() { pthread_mutex_lock(&mutex_arr); arrived++; if (arrived<N) { pthread_cond_wait(&cond, &mutex_arr); } else { pthread_cond_broadcast(&cond); arrived=0; /* be prepared for next barrier */ } pthread_mutex_unlock(&mutex_arr); }

29. Parallel SOR with Barriers (1 of 2) void* sor (void* arg) { int slice = (int)arg; int from = (slice * (n-1))/p + 1; int to = ((slice+1) * (n-1))/p + 1; for some number of iterations { … } }

30. Parallel SOR with Barriers (2 of 2) for (i=from; i<to; i++) for (j=1; j<n; j++) temp[i][j] = 0.25 * (grid[i-1][j] + grid[i+1][j] + grid[i][j-1] + grid[i][j+1]); barrier(); for (i=from; i<to; i++) for (j=1; j<n; j++) grid[i][j]=temp[i][j]; barrier();

31. Parallel SOR with Barriers: main int main(int argc, char *argv[]) { pthread_t *thrd[p]; /* Initialize mutex and condition variables */ for (i=0; i<p; i++) pthread_create (&thrd[i], &attr, sor, (void*)i); for (i=0; i<p; i++) pthread_join (thrd[i], NULL); /* Destroy mutex and condition variables */ }

32. Note again Many shared memory programming systems (other than Pthreads) have barriers as basic primitive. If they do, you should use it, not construct it yourself. Implementation may be more efficient than what you can do yourself.

33. Busy Waiting Not an explicit part of the API. Available in a general shared memory programming environment.

34. Busy Waiting initially: flag = 0; P1:produce data; flag = 1; P2:while( !flag ) ; consume data;

35. Use of Busy Waiting On the surface, simple and efficient. In general, not a recommended practice. Often leads to messy and unreadable code (blurs data/synchronization distinction). May be inefficient

36. Private Data in Pthreads To make a variable private in Pthreads, you need to make an array out of it. Index the array by thread identifier, which you should keep track of. Not very elegant or efficient.

37. Other Primitives in Pthreads Set the attributes of a thread. Set the attributes of a mutex lock. Set scheduling parameters.

38. ECE 1747 Parallel Programming Machine-independent Performance Optimization Techniques

39. Returning to Sequential vs. Parallel Sequential execution time: t seconds. Startup overhead of parallel execution: t_st seconds (depends on architecture) (Ideal) parallel execution time: t/p + t_st. If t/p + t_st > t, no gain.

40. General Idea Parallelism limited by dependences. Restructure code to eliminate or reduce dependences. Sometimes possible by compiler, but good to know how to do it by hand.

41. Optimizations: Example 16 for (i = 0; i < 100000; i++) a[i + 1000] = a[i] + 1; Cannot be parallelized as is. May be parallelized by applying certain code transformations.

44. Summary Reorganize code such that –dependences are removed or reduced –large pieces of parallel work emerge –loop bounds become known –…–… Code can become messy … there is a point of diminishing returns.

45. Factors that Determine Speedup Characteristics of parallel code –granularity –load balance –locality –communication and synchronization

46. Granularity Granularity = size of the program unit that is executed by a single processor. May be a single loop iteration, a set of loop iterations, etc. Fine granularity leads to: –(positive) ability to use lots of processors –(positive) finer-grain load balancing –(negative) increased overhead

47. Granularity and Critical Sections Small granularity => more processors => more critical section accesses => more contention.

48. Issues in Performance of Parallel Parts Granularity. Load balance. Locality. Synchronization and communication.

49. Load Balance Load imbalance = different in execution time between processors between barriers. Execution time may not be predictable. –Regular data parallel: yes. –Irregular data parallel or pipeline: perhaps. –Task queue: no.

51. Static vs. Dynamic Static: done once, by the programmer –block, cyclic, etc. –fine for regular data parallel Dynamic: done at runtime –task queue –fine for unpredictable execution times –usually high overhead Semi-static: done once, at run-time

52. Choice is not inherent MM or SOR could be done using task queues: put all iterations in a queue. –In heterogeneous environment. –In multitasked environment.

53. Static Load Balancing Block –best locality –possibly poor load balance Cyclic –better load balance –worse locality Block-cyclic –load balancing advantages of cyclic (mostly) –better locality

54. Dynamic Load Balancing (1 of 2) Centralized: single task queue. –Easy to program –Excellent load balance Distributed: task queue per processor. –Less communication/synchronization

55. Dynamic Load Balancing (2 of 2) Task stealing: –Processes normally remove and insert tasks from their own queue. –When queue is empty, remove task(s) from other queues. Extra overhead and programming difficulty. Better load balancing.

56. Semi-static Load Balancing Measure the cost of program parts. Use measurement to partition computation. Done once, done every iteration, done every n iterations.

57. Molecular Dynamics (MD) Simulation of a set of bodies under the influence of physical laws. Atoms, molecules, celestial bodies,... Have same basic structure.

58. Molecular Dynamics (Skeleton) for some number of timesteps { for all molecules i for all other molecules j force[i] += f( loc[i], loc[j] ); for all molecules i loc[i] = g( loc[i], force[i] ); }

59. Molecular Dynamics To reduce amount of computation, account for interaction only with nearby molecules.

60. Molecular Dynamics (continued) for some number of timesteps { for all molecules i for all nearby molecules j force[i] += f( loc[i], loc[j] ); for all molecules i loc[i] = g( loc[i], force[i] ); }

61. Molecular Dynamics (continued) for each molecule i number of nearby molecules count[i] array of indices of nearby molecules index[j] ( 0 <= j < count[i])

62. Molecular Dynamics (continued) for some number of timesteps { for( i=0; i<num_mol; i++ ) for( j=0; j<count[i]; j++ ) force[i] += f(loc[i],loc[index[j]]); for( i=0; i<num_mol; i++ ) loc[i] = g( loc[i], force[i] ); }

63. Molecular Dynamics (simple) for some number of timesteps { parallel for for( i=0; i<num_mol; i++ ) for( j=0; j<count[i]; j++ ) force[i] += f(loc[i],loc[index[j]]); parallel for for( i=0; i<num_mol; i++ ) loc[i] = g( loc[i], force[i] ); }

64. Molecular Dynamics (simple) Simple to program. Possibly poor load balance –block distribution of i iterations (molecules) –could lead to uneven neighbor distribution –cyclic does not help

65. Better Load Balance Assign iterations such that each processor has ~ the same number of neighbors. Array of “assign records” –size: number of processors –two elements: beginning i value (molecule) ending i value (molecule) Recompute partition periodically

66. Frequency of Balancing Every time neighbor list is recomputed. –once during initialization. –every iteration. –every n iterations. Extra overhead vs. better approximation and better load balance.

67. Summary Parallel code optimization –Critical section accesses. –Granularity. –Load balance.