程序代写代做代考 algorithm cache compiler 1

1

CS402: Seminar 2

An Introduction to OpenMP

OpenMP (Multiprocessing) is an API and runtime which enables the pro-

gramming of multiple processing cores with shared memory. The API is a

collection of functions and pragmas, the former allows the querying of informa-

tion, such as the number of active threads; and the latter allows the definition

of parallel regions. It is in these parallel regions, that threads are spawned by

the OpenMP runtime and the code within executed on multiple cores. These

threads are then able to communicate with each other using shared memory.

This lab will give a basic introduction in how to use the OpenMP API to exploit

parallelism in applications.

1 Hello World in OpenMP

As with the other labs, the first example we will look at is a variant of

“Hello, world!” which has been modified such that each thread prints out

its ID and “Hello, world!”. You can find the code in Listing 1 and the file

named helloWorldOMP.c.

#include

#include
#include

int main( int argc , char *argv [ ] ) {
/* D ef in e p ar a l l e l r e g io n * /
#pragma omp parallel

{
int thread_id ;

/* Use the OpenMP API to query wh ich thread we are in */

thread_id = omp_get_thread_num () ;

p r i n t f (” Hello World from thread: 91cd\n” , thread_id ) ;

/* En su r e a l l th r e ads h ave p r in ted be f o r e con t in u in g * /
#pragma omp barrier

/* Ac t ion on l y f r om th e th r ead wi th ID 0 * /
i f ( thread_ id == 0 ){

/ * P r in t th e t o t a l th e t o t a l n u m b e r o f th r e ad s * /

p r i n t f (“Total number of threads: 91cd\n” ,
omp_get_num_threads () ) ;

}

}

return 0;

}

Listing 1: OpenMP “Hello World” example

2

This example can be compiled with the following command: gcc

-fopenmp -o helloWorldOMP.b helloWorldOMP.c

The only difference from the normal compilation process is the addition of the

-fopenmp flag. This is a compiler flag rather than just a collection of libraries,

because OpenMP requires compiler support for the additional language features

it provides. Run the compiled binary just like you would any other application,

and you will be greeted with output similar to that below:

Hello World from thread: 0

Hello World from thread: 1

Total number of threads: 2

Your output may vary from the above for two reasons: 1) there is no guar-

antee to the order in which each thread will print “Hello, world!” and; 2) by

default the OpenMP run time will choose the number of threads to spawn in a

parallel region based on your hardware. In this case the runtime has selected 2

threads as this is what it thinks is best for the available hardware. It is possible

to override this decision by setting the OMP_NUM_THREADS environmental

variable. For this lab we will leave the OpenMP runtime to decide how many

threads should be used. Congratulations, you have compiled and run a basic

OpenMP program, now we will look at each part of the program in more detail.

 #include – Necessary for accessing the functions provided by the
OpenMP runtime, such as omp_get_thread_num().

 #pragma omp parallel – Define the following region (the area scoped
by { and }) as to be executed in parallel. If this pragma is used without

braces then it will default to the line of code below it, so in a similar manner

to how the scope of if statements operate.

 omp_get_thread_num() – A function which returns the ID of
the thread the program is executing in.

 #pragma omp barrier – Instruct the OpenMP runtime to wait for all
threads to reach this point before continuing. This is how we can force all

threads to print “Hello, world!” before thread 0 reports the total number of

threads.

 omp_get_num_threads() – A function which reports the number of
threads which the OpenMP runtime has spawned in a parallel region.

As always, if you have any questions at this point, make sure you ask

one of the tutors for help.

2 Parallel Loops

Defining parallel regions with #pragma omp parallel is just one method

by which OpenMP allows parallelism to be exploited; another mechanism which

OpenMP provides allows iterations of a loop to be scheduled across available

threads. This is done by proceeding a for loop with the following pragma:

3

#pragma omp paral lel for

There are several options to this pragma, including those which allow the

selection of scheduling algorithm and the scoping of variables. These two options will

be detailed here, but it is recommended to review them all by reading the appropriate

sections of the OpenMP specification.’.

 schedule (static) – Iterations are divided up into chunks, and these

consecutive groups of iterations are spread among the threads using a

round robin policy. This method tends to be good when you want good

cache access patterns.

 schedule (dynamic) – Iterations are again split into chunks, but when a

thread finished a chunk of it will request the next chunk. In this way you

can load balance if your iterations take varying amounts of time to

complete.

 private([list_of_variable_names,…]) – It is common to require

variables which are local to a loop, such as the loop counter and any

partial results.

The remainder of this lab will focus on the code in Listing 2, which is a

serial implementation of integration using the Trapezoidal Rule. We will step

through the parallelisation of this code in the tasks which follow.

#include

#include

float f ( float x) ;

int main ( int argc , char** argv ) {
int a = 0 . 0 ; // Left endpoint
int b = 1 . 0 ; // Right endpoint

long n = 1024000000; // Number of trapezoids
double h = (b − a) / ((double) n) ; // Base length
double integral = ( f (a) + f (b) ) / 2 . 0 ;
long i ;

double x = a ;

for ( i = 1 ; i <= n ; i ++) { x = x + h ; i n t e g r a l = i n t e g r a l + f ( x ) ; } integral = i n t e g r a l *h ; p r i n t f ( "With n = %d trapezoids , estimate : %f \n" , n , integral) ; } // The function to integrate , here we use 2 / ( x^4 + 1) float f ( float x) { return 2 / ( ( x*x*x*x) + 1) ; } Listing 2: Integration using the Trapezoidal Rule 1 http://openmp.org/ 4 Task 1 Compile the code in Listing 2 as you have done with the previous examples, execute the resultant binary, and record the values of the following: a, b, n and the computed area. Additionally, it is useful to be able to record the amount of time a piece of code has taken to execute, such as for comparison against other versions of the same code. OpenMP conveniently provides the following function:  double omp_get_wtime() – This function returns the time elapse in seconds from some fixed point (which is consistent for multiple calls in a program run). Using omp_get_wtime() instrument the code in Listing 2 such that it re- ports the time taken to execute the for loop and the following operation on the integral variable. Once this is done, execute the program and note down the runtime; you may need to adjust the value of n such that the code takes several seconds to complete. It may also be worth modifying the program output to report the values of a, b and n in order to make recording easier. Task 2 Now we have a way to easily compare runs based on the input param- eters, runtime and the result, we can begin parallelising this code. The obvious place to start is the for loop as this is where all the computation is done. Place the following pragma above the for loop:  #pragma omp parallel for private(i) schedule(static) – Instruct the OpenMP runtime to schedule iterations of the loop statically between the available threads. Additionally, specify that the loop counter i is private, so that each thread has its own copy. With this pragma in place, we must now consider how to modify the loop body to cope with iterations being scheduled. That is, we can no longer assume that iterations will be run in any kind of order, therefore to compute the value of x incrementally is invalid. Your next action for this task is to modify the computation of x to be in terms of a, h and i. Compile and run the code, then compare the result and the runtime to that of the serial version. Task 3 As you can see, the addition of the OpenMP pragma ruined the result. This is because the program is summing into the integral from all threads - it is an unguarded critical section. Place the following pragma above the integral summation (integral = integral + f(x)):  #pragma omp critical – Instruct the OpenMP runtime to only allow a single thread to run this code at any given time. Again, compile and run the code, and compare the result and runtime to that of the serial code. You will probably notice that while the result is now the same, the performance of the application has been greatly affected by the critical section. 5 3 Reductions Luckily the computational pattern of summing between threads is a common one, so there is support for this in the OpenMP runtime. This type of operation is know as a reduction and is specified as another option to #pragma omp parallel for.  reduction(operator:variable) – The reduction operator takes an argument of the form operator:variable where operator is for example +,-,/,* and variable is the identifier of the variable you wish to reduce over. Task 4 Using the knowledge of reductions which you have just gained, save your code to Task 3 in a new file and modify it to use reductions rather than critical sections. Compile and run the code, and then compare the result and runtime to that of the code from Task 3. Task 5 (Optional) Using what you have learn in the lab, write a parallel implementation of the Trapezoidal Rule using OpenMP which avoids the use of a reduction every iteration. Compare this runtime of this against your imple- mentation from Task 4. References [1] Peter Pacheco, An Introduction to Parallel Programming, 2011.