Your SlideShare is downloading. ×
0
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×
Saving this for later? Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime – even offline.
Text the download link to your phone
Standard text messaging rates apply

Algoritmi e Calcolo Parallelo 2012/2013 - OpenMP

266

Published on

Slide del corso di Algoritmi e Calcolo Parallelo per il corso di laurea magistrale in Ingegneria Matematica 2012/2013 - Politecnico di Milano

Slide del corso di Algoritmi e Calcolo Parallelo per il corso di laurea magistrale in Ingegneria Matematica 2012/2013 - Politecnico di Milano

Published in: Education, Technology
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
266
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
8
Comments
0
Likes
0
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. OpenMP Algoritmi e Calcolo Parallelo Prof. Pier Luca Lanzi
  • 2. References • • • 2 Using OpenMP: Portable Shared Memory Parallel Programming, Barbara Chapman, Gabriele Jost and Ruud van der Pas OpenMP.org http://openmp.org/ OpenMP Tutorial https://computing.llnl.gov/tutorials/openMP/ Prof. Pier Luca Lanzi
  • 3. Prologue Prof. Pier Luca Lanzi
  • 4. Threads for Dummies (1) • • • 4 A thread is a sequence of instructions to be executed within a program Usually a program consist of a single thread of execution that starts in main(). Each line of your code is executed in turn, exactly one line at a time. A (UNIX) process consists of A virtual address space (i.e., stack, data, and code segments) System resources (e.g., open files) A thread of execution    Prof. Pier Luca Lanzi
  • 5. Threads for Dummies (2) • • • • 5 In a threaded environment, one process consists of multiple threads Each user process can have several threads of execution, different parts of the user’s code can be executing simultaneously The collection of threads form a single process! Each thread in the process sees the same virtual address space, files, etc. Prof. Pier Luca Lanzi
  • 6. You can manage the multiple threads in your process yourself (but it is quite complex!) Or you can let OpenMP do it almost automatically! Prof. Pier Luca Lanzi
  • 7. The Parallelization Model of OpenMP Prof. Pier Luca Lanzi 7
  • 8. Fork Join Parallelization Prof. Pier Luca Lanzi 8
  • 9. Introduction Prof. Pier Luca Lanzi
  • 10. OpenMP = Open Multi-Processing API for multi-platform shared memory multiprocessing programming Supports C/C++ and Fortran Available on Unix and on MS Windows Compiler directives, Library routines, & Environment variables Prof. Pier Luca Lanzi
  • 11. History of OpenMP • • • • 11 The OpenMP Architecture Review Board (ARB) published its first API specifications, OpenMP for Fortran 1.0, in October 1997. October the following year they released the C/C++ standard. 2000 saw version 2.0 of the Fortran specifications with version 2.0 of the C/C++ specifications being released in 2002. Version 2.5 is a combined C/C++/Fortran specification that was released in 2005. Version 3.0, released in May, 2008, is the current version of the API specifications. Included in the new features in 3.0 is the concept of tasks and the task construct. These new features are summarized in Appendix F of the OpenMP 3.0 specifications. Prof. Pier Luca Lanzi
  • 12. What are the Goals of OpenMP? • • • Standardization Provide a standard among a variety of shared memory architectures Lean and Mean Establish a simple and limited set of directives for programming shared memory machines (significant parallelism can be implemented by using just 3 or 4 directives) Ease of Use Provide capability to incrementally parallelize a serial program (unlike message-passing libraries) Provide the capability to implement both coarse-grain and fine-grain parallelism Portability Supports Fortran (77, 90, and 95), C, and C++     • 12  Prof. Pier Luca Lanzi
  • 13. How Does It Work? • • • • 13 OpenMP is an implementation of multithreading A master thread "forks" a specified number of slave threads Tasks are divided among slaves Slaves run concurrently as the runtime environment allocating threads to different processors Prof. Pier Luca Lanzi
  • 14. Most OpenMP constructs are compiler directives or pragmas The focus of OpenMP is to parallelize loops OpenMP offers an incremental approach to parallelism Prof. Pier Luca Lanzi
  • 15. Overview of OpenMP • 15 A compiler directive in C/C++ is called a pragma (pragmatic information). It is a preprocessor directive, thus it is declared with a hash (#). Compiler directives specific to OpenMP in C/C++ are written in codes as follows: #pragma omp <rest of pragma> Prof. Pier Luca Lanzi
  • 16. The OpenMP Programming Model • • • • • 16 Based on compiler directives Nested Parallelism Support API allows parallel constructs inside other parallel constructs Dynamic Threads API allows to dynamically change the number of threads which may used to execute different parallel regions Input/Output OpenMP specifies nothing about parallel I/O The programmer has to insure that I/O is conducted correctly within the context of a multi-threaded program Memory consistency Threads can "cache" their data and are not required to maintain exact consistency with real memory all of the time When it is critical that all threads view a shared variable identically, the programmer is responsible for insuring that the variable is FLUSHed by all threads as needed       Prof. Pier Luca Lanzi
  • 17. Basic Elements Prof. Pier Luca Lanzi
  • 18. Hello World 18 #include <iostream> using namespace std; int main() { cout << "Hello Worldn"; } $ ./hello Hello World Prof. Pier Luca Lanzi
  • 19. Hello World in OpenMP #include <omp.h> $ ./hello #include <iostream> Hello World using namespace std; int main() Hello World { Hello World #pragma omp parallel num_threads(3) { cout << "Hello Worldn”; // what if // cout << "Hello World" << endl; } } Prof. Pier Luca Lanzi 19
  • 20. Compiling an OpenMP Program • • 20 OpenMP consists of a set of directives and macros that are translated in C/C++ and thread instructions To compile a program using OpenMP, we need to include the flag “openmp”, g++ -o example ./source/openmp00.cpp –fopenmp • The command compile the file openmp00.cpp using openmp and generates the executable named “example” Prof. Pier Luca Lanzi
  • 21. What Does Really Happen? • • 21 OpenMP consists of a set of directives and macros that are translated in C/C++ and thread instructions We can check how the original program using OpenMP is translated, using the command g++ -fdump-tree-omplower ./source/openmp00.cpp – fopenmp Prof. Pier Luca Lanzi
  • 22. OpenMP Directives • • • • A valid OpenMP directive must appear after the #pragma omp name name identify an OpenMP directive The directive name can be followed by optional clauses (in any order) An OpenMP directive precedes the structured block which is enclosed by the directive itself #pragma omp name [clause, ...] { … } • 22 Example #pragma omp parallel num_threads(3) { cout << "Hello Worldn"; } Prof. Pier Luca Lanzi
  • 23. Directive Scoping • 23 Static (Lexical) Extent The code textually enclosed between the beginning and the end of a structured block following a directive The static extent of a directives does not span multiple routines or code files Orphaned Directive An OpenMP directive that appears independently from another enclosing directive. It exists outside of another directive's static (lexical) extent. Will span routines and possibly code files Dynamic Extent The dynamic extent of a directive includes both its static (lexical) extent and the extents of its orphaned directives.   • •    Prof. Pier Luca Lanzi
  • 24. Directive Scoping: example void f(){ #pragma omp name2 [clause, ...] { … } } int main(){ #pragma omp name1 [clause, ...] { f(); … } } Prof. Pier Luca Lanzi 24
  • 25. Directive Scoping: example 25 void f(){ #pragma omp name2 [clause, ...] { … } } Static extent of name1 int main(){ #pragma omp name1 [clause, ...] { f(); … } } Prof. Pier Luca Lanzi
  • 26. Directive Scoping: example void f(){ #pragma omp name2 [clause, ...] { … } Orphaned directive } Dynamic extent of name1 int main(){ #pragma omp name1 [clause, ...] { f(); … } } Prof. Pier Luca Lanzi
  • 27. Parallel Construct Prof. Pier Luca Lanzi
  • 28. The Parallel Construct/Directive • 28 Plays a crucial role in OpenMP, a program without a parallel construct will be executed sequentially #pragma omp parallel [clause, clause, ... ] { // block } • • When a thread reaches a parallel directive, it creates a team of threads and becomes the master of the team. The master is a member of that team and has thread number 0 within that team. Prof. Pier Luca Lanzi
  • 29. The Parallel Construct/Directive • • • • 29 Starting from the beginning of the parallel region, the code is duplicated and all threads will execute that code. There is an implied barrier at the end of a parallel section. Only the master thread continues execution past this point. If any thread terminates within a parallel region, all threads in the team will terminate, and the work done up until that point is undefined. Threads are numbered from 0 (master thread) to N-1 Prof. Pier Luca Lanzi
  • 30. How Many Threads? • 30 The number of threads in a parallel region is determined by the following factors, in order of precedence: Evaluation of the if clause Setting of the num_threads clause Use of the omp_set_num_threads() library function Setting of the OMP_NUM_THREADS environment variable Implementation default, e.g., the number of CPUs on a node      • When present, the if clause must evaluate to true (i.e., non-zero integer) in order for a team of threads to be created; otherwise, the region is executed serially by the master thread. Prof. Pier Luca Lanzi
  • 31. Run-Time Library Routines Prof. Pier Luca Lanzi
  • 32. Overview • 32 The OpenMP standard defines an API for a variety of library functions: Query the number of threads/processors, set number of threads to use General purpose locking routines (semaphores) Portable wall clock timing routines    • • • Set execution environment functions: nested parallelism, dynamic adjustment of threads It may be necessary to specify the include file "omp.h". For the lock routines/functions The lock variable must be accessed only through the locking routines The lock variable must have type omp_lock_t or type omp_nest_lock_t, depending on the function being used.   Prof. Pier Luca Lanzi
  • 33. Run-Time Routines • 33 void omp_set_num_threads(int num_threads)  Sets the number of threads that will be used in the next parallel region num_threads must be a postive integer   This routine can only be called from the serial portions of the code • int omp_get_num_threads(void)  Returns the number of threads that are currently in the team executing the parallel region from which it is called • int omp_get_thread_num(void)  Returns the thread number of the thread, within the team,  making this call. This number will be between 0 and OMP_GET_NUM_THREADS1. The master thread of the team is thread 0 Prof. Pier Luca Lanzi
  • 34. A better Hello World in OpenMP 34 #include <omp.h> #include <iostream> using namespace std; int main() { #pragma omp parallel num_threads(3) { cout << "Hello World. I am thread"; cout << omp_get_thread_num() << endl; } } $ ./hello Hello World. I am thread 0 Hello World. I am thread 2 Hello World. I am thread 1 Prof. Pier Luca Lanzi
  • 35. Conditional Compiling // openmp03.cpp #ifdef _OPENMP #include <omp.h> #else #define omp_get_thread_num() 0 #endif ... // numero del thread corrente int TID = omp_get_thread_num(); ... Prof. Pier Luca Lanzi 35
  • 36. Data Scope Attribute Clauses Prof. Pier Luca Lanzi
  • 37. Variables Scope in OpenMP • • • 37 OpenMP is based upon the shared memory programming model, most variables are shared by default The OpenMP Data Scope Attribute Clauses are used to explicitly define how variables should be scoped Four main scope types private shared default reduction     Prof. Pier Luca Lanzi
  • 38. Variables Scope in OpenMP • 38 Data Scope Attribute Clauses are used in conjunction with several directives to Define how and which data variables in the serial section of the program are transferred to the parallel sections Define which variables will be visible to all threads in the parallel sections and which variables will be privately allocated to all threads   • Data Scope Attribute Clauses are effective only within their lexical/static extent. Prof. Pier Luca Lanzi
  • 39. The private Clause • • • 39 Syntax: private (list) The private clause declares variables in its list to be private to each thread Private variables behave as follows: a new object of the same type is declared once for each thread in the team all references to the original object are replaced with references to the new object variables should be assumed to be uninitialized for each thread    Prof. Pier Luca Lanzi
  • 40. Example: private Clause 40 #pragma omp parallel for private(i,a) for (i=0; i<n; i++) { a = i+1; printf("Thread %d has a value of a = %d for i = %dn", omp_get_thread_num(),a,i); } /*-- End of parallel for --*/ Prof. Pier Luca Lanzi
  • 41. The shared Clause • • • • 41 Syntax: shared (list) The shared clause declares variables in its list to be shared among all threads in the team. A shared variable exists in only one memory location and all threads can read or write to that address. It is the programmer's responsibility to ensure that multiple threads properly access shared variables Prof. Pier Luca Lanzi
  • 42. Example: shared Clause #pragma omp parallel for shared(a) for (i=0; i<n; i++) { a[i] += i; } /*-- End of parallel for --*/ Prof. Pier Luca Lanzi 42
  • 43. The default clause • • • • • 43 Syntax: default (shared | none) The default scope of all variables is either set to shared or none The default clause allows the user to specify a default scope for all variables in the lexical extent of any parallel region Specific variables can be exempted from the default using specific clauses (e.g., private, shared, etc.) Using none as a default requires that the programmer explicitly scope all variables. Prof. Pier Luca Lanzi
  • 44. The reduction Clause • • • • • 44 Syntax: reduction (operator: list) Performs a reduction on the variables that appear in its list Operator defines the reduction operation A private copy for each list variable is created for each thread At the end of the reduction, the reduction variable is applied to all private copies of the shared variable, and the final result is written to the global shared variable Prof. Pier Luca Lanzi
  • 45. Work-Sharing Prof. Pier Luca Lanzi
  • 46. Work-Sharing Constructs • • • • 46 Divide the execution of the enclosed code region among the members of the team that encounter it A work-sharing construct must be enclosed dynamically within a parallel region in order for the directive to execute in parallel Work-sharing constructs do not launch new threads There is no implied barrier upon entry to a work-sharing construct, however there is an implied barrier at the end of a work sharing construct for: shares iterations of a loop across the team (data parallelism) sections: breaks work into separate, discrete sections, each executed by a thread (functional parallelism) Prof. Pier Luca Lanzi single: serializes a section of code.
  • 47. The for Construct/Directive • • • • • 47 #pragma omp for [clause ...] for_loop Three main clause types: schedule, nowait, and the data scope schedule (type [,chunk]) describes how iterations of the loop are divided among the threads in the team (default schedule is implementation dependent) nowait if specified, then threads do not synchronize at the end of the parallel loop. The data-scope clauses (private, shared, reduction) Prof. Pier Luca Lanzi
  • 48. The for Construct/Directive (Schedule Types) • • • • 48 static: loop iterations are divided into blocks of size chunk and then statically assigned to threads If chunk is not specified, the iterations are evenly (if possible) divided contiguously among the threads dynamic: loop iterations are divided into blocks of size chunk, and dynamically scheduled among the threads When a thread finishes one chunk, it is dynamically assigned another. The default chunk size is 1 Prof. Pier Luca Lanzi
  • 49. Example: Dot Product (1) int main() { double sum; vector<double> a(10000000), b(10000000); generate(a.begin(),a.end(),drand48); generate(b.begin(),b.end(),drand48); sum = 0; for (int i=0; i<a.size(); ++i) { sum = sum + a[i]*b[i]; } cout << sum << endl; } Prof. Pier Luca Lanzi 49
  • 50. Example: Dot Product (2) // openmp03-dotp-openmp int main() { double sum; vector<double> a(10000000), b(10000000); generate(a.begin(),a.end(),drand48); generate(b.begin(),b.end(),drand48); sum = 0; #pragma omp parallel for reduction(+: sum) for (int i=0; i<a.size(); ++i) { sum = sum + a[i]*b[i]; } cout << sum << endl; } Prof. Pier Luca Lanzi 50
  • 51. Example: Dot Product (3) 51 // openmp03-dotp-openmp3 ... int ChunkSize = a.size()/omp_get_max_threads(); #pragma omp parallel { #pragma omp for schedule(dynamic, ChunkSize) reduction(+: sum) for (int i=0; i<a.size(); ++i) { sum = sum + a[i]*b[i]; } } Prof. Pier Luca Lanzi
  • 52. Example: Varying Loads 52 // openmp06 ... void f(vector<double> &a, vector<double> &b, int n) { int i, j; #pragma omp parallel shared(n) { // schedule is dynamic of size 1 to manage the varying size of load #pragma omp for schedule(dynamic,1) private (i,j) nowait for (i = 1; i < n; i++) for (j = 0; j < i; j++) b[j + n*i] = (a[j + n*i] + a[j + n*(i-1)]) / 2.0; } } // // // // time ./openmp04-seq 10000 (n is 10000^2) 11.20 real 5.68 user 0.68 sys time ./openmp04 10000 4.55 real 13.07 user 0.62 sys Prof. Pier Luca Lanzi
  • 53. The sections Construct/Directive #pragma omp sections [clause ...] { #pragma omp section { // execute A } #pragma omp section { // execute A } } • • 53 private reduction nowait etc. Specifies that the enclosed section(s) of code are to be divided among the threads in the team Each section is executed once by a thread in the team. Different sections may be executed by different threads. It is possible that for a thread to execute more than one section. Prof. Pier Luca Lanzi
  • 54. Example: Dot Product (vectors init) #pragma omp sections nowait { #pragma omp section { int ita; #pragma omp parallel for shared(a) private(ita) schedule(dynamic,ChunkSize) for(ita=0; ita<a.size(); ++ita) { a[ita] = drand48(); } } #pragma omp section { int itb; #pragma omp parallel for shared(b) private(itb) schedule(dynamic,ChunkSize) for(itb=0; itb<b.size(); ++itb) { b[itb] = drand48(); } } } Prof. Pier Luca Lanzi 54
  • 55. The single Construct/Directive • • • 55 The single construct is associated with the structured block of code immediately following it and specifies that this block should be executed by one thread only It does not state which thread should execute the code block The thread chosen could vary from one run to another. It can also differ for different single constructs within one application #pragma omp single [private | nowait | …] { ... Prof. Pier Luca Lanzi }
  • 56. Example: Single Construct/Directive #pragma omp parallel shared(a,b) private(i) { #pragma omp single { a = 10; printf("Single construct executed by thread %dn", omp_get_thread_num()); } /* A barrier is automatically inserted here */ #pragma omp for for (i=0; i<n; i++) b[i] = a; } /*-- End of parallel region --*/ Prof. Pier Luca Lanzi 56
  • 57. Synchronization Prof. Pier Luca Lanzi
  • 58. The critical Construct/Directive • • • • 58 #pragma omp critical [ name ] { … } The critical directive specifies a region of code that must be executed by only one thread at a time If a thread is currently executing inside a critical region and another thread reaches that critical region and attempts to execute it, it will block until the first thread exits that critical region The optional name enables multiple different critical regions Names act as global identifiers Critical regions with the same name are treated as the same region All critical sections which are unnamed, are treated as the same section    Prof. Pier Luca Lanzi
  • 59. Example: critical Construct/Directive 59 sum = 0; #pragma omp parallel shared(n,a,sum) private(TID,sumLocal) { TID = omp_get_thread_num(); double sumLocal = 0; #pragma omp for for (size_t i=0; i<n; i++) sumLocal += a[i]; #pragma omp critical (update_sum) { sum += sumLocal; cout << "TID=" << TID << "sumLocal=" << sumLocal << " sum=" << sum << endl; } } /*-- End of parallel region --*/ cout << "Value of sum after parallel region: " << sum << endl; Prof. Pier Luca Lanzi
  • 60. The barrier Construct/Directive • • • • 60 Syntax: #pragma omp barrier The barrier directive synchronizes all threads in the team. When a barrier directive is reached, a thread will wait at that point until all other threads have reached that barrier. All threads then resume executing in parallel the code that follows the barrier. All threads in a team (or none) must execute the barrier region. Prof. Pier Luca Lanzi
  • 61. When to Use OpenMP? Target platform is multicore or multiprocessor Application is cross-platform Parallelizable loops Last-minute optimizations Prof. Pier Luca Lanzi

×