OpenMP

'Parallel: Using more than one processor at a time'

=Introduction=

Computers. They're handy right? Email, buying stuff from Amazon.. You could even write a program. Create your own application. Indeed, for people of a certain mindset--people like you and me--we can do all sorts of interesting things like simulating the natural world, and in the process look at questions like, "will Greenland melt if we carry on polluting the atmosphere as we are?" and "what will be the consequences if it does melt?".

Sometimes it's handy to have more than one computer. Let's say that we have a new whizz-bang weather model that takes 26 hours to work out what the weather will do tomorrow. "Very nice. Superb accuracy," you say, "but, alas, about as much use as a chocolate teapot." In order for the model to be useful in the real world, we need it to run faster. We need to divide up the work it does and run it over two, or more computers. We need to enter the world of parallel programming.

"Hippee!" we cry, but a word of caution. Getting models to work reliably in parallel is a lot--I'll say it again, 'a lot--harder than getting them to work on a single processor. Before setting out down the road, it is well worth checking that you really do need your model to run faster, and that you've explored all other avenues to speed up your model before embarking on a parallelisation project.

Still keen? OK, let's get stuck in.

=A Necessary Aside=

We'll look at some parallel code in a moment--I promise--but before we do we must take a look at some simple serial code and ask ourselves whether we've been as smart as we can and made the code run as fast as it can. I know you're keen to crack on, but trust me, it's really important to make sure that your serial code optimised before converting it into parallel code. If you don't, any inefficiencies in your code are just going to get magnified as it is run on several processors at the same time, and that's not what you want.

Let's get some examples involving 2-dimensional arrays and nested loops from another short course on how to profile your code:

svn co http://source.ggy.bris.ac.uk/subversion-open/profiling/trunk ./profiling cd profiling/examples/example2 make

and follow the accompanying notes.

Setting up your loops the wrong way round is bad enough for serial code. If you were to incorporate those same loops in some parallel code, you may very well end up creating a so-called false sharing condition and your parallel code will end up running very slowly indeed.

Righteo, forewarned is forearmed. We know that before we embark on any code parallelisation, we need to profile our code and ged rid of any inefficiencies. Once we've done that, we're ready to start the task of making it run in parallel. Indeed, if you are lucky, you may find that your newly optimised code has been sped-up sufficiently and you don't need to do anything else to it!

=OpenMP=

There are a number of different ways to create parallel programs, and we're going to start with one approach, called OpenMP. There a number of reasons for this:


 * 1) It's widely available,
 * 2) It's good for the muli-core processors that we find in a lot of computers these days,
 * 3) It's fairly easy to use,
 * 4) and it's based upon the not-so-minde-bending concept of threads



At this point, we could launch overselves into a long and detailed discussion of threads, the OpenMP runtime environment, pre-processor macro statements and the like. But we won't. Because it's less fun. Let's just try an example instead.

OK, to get the examples, login to a Linux box and cut & paste the below onto the command line:

svn co http://source.ggy.bris.ac.uk/subversion-open/parallel/trunk ./parallel

=Hello, world=

Once you have the examples, do the following:

cd examples/example1 make ./omp_hello_f90.exe

Tada! Just like the old classic 'hello, world', but this time run in parallel on as many processors as you have available on your machine. Good eh? Now, how did we do that?

Take a look inside the file omp_hello_f90.f90. First up we have used a Fortran90 module containing routines specific to OpenMP:

use omp_lib

This gives us access to routines like:

omp_get_num_threads

and

omp_get_thread_num

The rest of the program is straightforward Fortran code, except for some comment lines starting with !$omp, such as:

!$omp parallel private(nthreads, tid) ... !$omp end parallel

These lines specify that the master thread should fork a parallel region. From the code, you will see that all the threads on the team will get their thread ID--through calls to the OpenMP library--and will print it. The master thread will also ask how many threads have been spawned by the OpenMP runtime environment, and will print the total.

Notice that the variables nthreads and tid have been marked as private. This means that seperate copies of the these variables will be kept for each thread. This is essential, or else the print statement, 'Hello, world from thread = ' would get all mixed up, right? Try deliberatly mucking things up. Go on, see what happens if you delete tid from the private list. You can always use, e.g.:

svn revert omp_hello_f90.f90

to get the original version of the program back, if you get into a pickle.

The line:

!$omp barrier

Specifies that all threads must reach this point before going on. A kind of catch-up point.

Look inside the Makefile and notice that the use of OpenMP has been flagged to the compiler. -fopenmp in this case, as we are using the GNU compiler, gfortran. (It would be just -openmp if you were using ifort, say.) If you were to try to compile the code without this flag, you'd get a compile-time error.

If you prefer C, there's a version of the program for you in omp_hello_c.c. All of the general comments above hold for this example too. You'll notice the use of:


 * 1) include 

(instead of use omp_lib)

and statements of the form:


 * 1) pragma omp ...

for the OpenMP instructions.

=Work Sharing inside Loops=

Great. We can run some code in parallel and it wasn't too tricky, right? As fun as 'hello, world' always is, the real benefits appear when we can divide up a task between different processors and let them get on with some real work in parallel. That way the program will run faster overall. In OpenMP-speak, this is worksharing. Let's look at a straightforward example:

cd ../example2

Again we have C and Fortran90 versions of the same program. Take a look inside simple_f90.f90. Here we see that we have a parallel section and each thread enquires of it's ID and the master enquires of the total number of threads. All as per example1. Notice this time, however, that some variables are marked as shared. In particular, the array a. Inside the parallel section, we have a do loop preceeded by the comment:

!$omp do

This statement is the key to the worksharing. It means that iterations from the do loop will be farmed out to different processors and performed in parallel. The easiest way to see what is going on is to run the program:

make ./simple_f90.f90

We can see that the iterations from the loop were chunked and that one processor placed values into a block of cells in the array (a) and that another processor was responsible for another block.

In addition to this default behaviour, OpenMP provides us with a number of keywords so that we may exert some finer grain control on the worksharing. To see an example of this take a look inside 'schedule_f90.f90. Here we see:

!$omp do schedule(static,chunk)

and that we have previously set the value of chunk to 10. The static keyword indicates that each chunk of 10 contiguous iterations will be farmed out to each thread in turn--round-robin fashion. Since we are initialising the values in the array in the loop, all threads need access to the same array (a) and that's why it is marked as shared.

There are a number of different scheduling keywords we can use. Try replacing static with dynamic. In this case chunks are handed out to processors as they become available. You may see a certain (quiet) processor get used several times and so all the processors may not be called upon in the workshare. Try changing the chunk size too. Go on, make a mess! It's the best way to learn:)

=Reductions=

A reduction combines elements from the different threads to form a single result. This is something that we might like to do quite often. For example, reductions can be summations or products. Let's look at the next example:

cd ../example3

Here we'll look at an iterative algorithm to compute a value for pi and we'll see how we can speed this up by spreading the iteration over several processors. As normal, there are C and Fortran90 examples that we can look at.

Note that in both the reduction_ programs we explicitly control the number of threads that we want OpenMP to use (set to 4 initially) with a call to:

omp_set_num_threads(NUM_THREADS)

(A function call in C and a subroutine call in Fortran.)

In the C code, we flag the parallel reduction using the pragma statement:


 * 1) pragma omp parallel for reduction(+:sum) private(x)

We have a similar comment line in the Fortran code. Note in this case that the parallel do statement has a corresponding end parallel do:

!$omp parallel do reduction(+:sum) private(x) ... !$omp end parallel do

The C-code example has only the:


 * 1) pragma omp parallel for ...

since the for block is structured and hence the end of the block is given by the syntax.

In both cases we see that sum is indicated as the reduction variable, and that x is kept private to each thread.

Now compile and run the serial and parallel programs, timing them as they go:

make time ./serial_pi_f.exe time ./reduction_pi_f.exe

Note how much quicker the parallel code runs. You could try changing the number of threads that we utilise (probably up to a maximum of 8, since the hardware most likely has a maximum of 8 processor-cores). Recompile using make, and re-time the runs.

=Matrix Multiplication=

The previous examples have been useful in highlighting the key features that OpenMP offers, but have been very simple. Let's move onto a more realistic task. Matrix multiplication is a good candidate for parallelisation:

cd ../example4

Again, let's compile the code and time the serial and parallel versions:

make time ./serial_mm_fort.exe time ./omp_mm_fort.exe

If you look inside serial_mm_fort.f90 you will see that after the initialisations have been done, the work of the matrix multiplication is done by three nested loops:

do ii=1, nra do jj=1, ncb do kk=1, nca c(ii,jj) = c(ii,jj) + a(ii,kk) * b(kk,jj) end do    end do  end do

In the OpenMP version, this is modified to:

!$omp do schedule(static, chunk) do ii=1, nra print *, 'thread', tid, 'did row', ii    do jj=1, ncb do kk=1, nca c(ii,jj) = c(ii,jj) + a(ii,kk) * b(kk,jj) end do    end do  end do

If you look carefully at the output, you will see that each processor does a chunk (set to 10) of the iterations of the outer loop (which has the index ii). The inner loops are done by the same processor.

Homework: Given our knowledge of row-major order and column-major order storage of 2-dimensional arrays, form the first section above, is there a better way to arrange the loops, and hence worksharing, for the C and Fortran examples?

=Some More Features=

In this section, I'll list some more useful features offered by OpneMP, but without example code to go with them, since I reckon you're getting the hang of things now:) I'll use the Fortran syntax for my descriptions, but the corresponding C constructs are as you would expect them to be.

Only one thread in the team will execute the enclosed block of code:

!$omp single ... !$omp end single

Only the master thread will execute the enclosed code:

!$omp master ... !$omp end master

Only one thread at a time will execute the block (for 'non-thread-safe' code):

!$omp critical ... !$omp end critical

An interesting construct that will ensure that the loop iterations are done in the same order as would be done by the corresponding piece of serial code:

!$omp do ordered ...loop region... !$omp ordered ... !$omp end ordered ...end of loop region... !$omp end do

=To Go Further=

A number of tutorials and example programs using OpenMP can be found by searching the Web. A few good ones that I've seen include:


 * A comprehensive tutorial from Lawrence Livermore National Lab
 * Comparative serial and OpenMP programs progressing to, e.g. one to solve the heat diffusion equation