Explore
For Enterprise
Join for Free
Log In

Artes e Humanas
Negócios
Ciência da Computação
Ciência de Dados
Tecnologia da informação
da saúdesaúde
Matemática e Lógica
Desenvolvimento Pessoal
Ciência e Engenharia Física
Ciências Sociais
Línguas
Notas
Certificados

Explore todo o Coursera

Navegar
Pesquisar
For Enterprise
Entrar
Inscreva-se gratuitamente

3.4 Parallel One-Dimensional Iterative Averaging

Loading...

Parallel Programming in Java

Universidade Rice

4.6 (619 classificações) | 17K alunos inscritos

Curso 1 de 3 no Parallel, Concurrent, and Distributed Programming in Java Programa de cursos integrados

This course teaches learners (industry professionals and students) the fundamental concepts of parallel programming in the context of Java 8. Parallel programming enables developers to use multicore computers to make their applications run faster by using multiple processors at the same time. By the end of this course, you will learn how to use popular parallel Java frameworks (such as ForkJoin, Stream, and Phaser) to write parallel programs for a wide range of multicore platforms including servers, desktops, or mobile devices, while also learning about their theoretical foundations including computation graphs, ideal parallelism, parallel speedup, Amdahl's Law, data races, and determinism. Why take this course? • All computers are multicore computers, so it is important for you to learn how to extend your knowledge of sequential Java programming to multicore parallelism. • Java 7 and Java 8 have introduced new frameworks for parallelism (ForkJoin, Stream) that have significantly changed the paradigms for parallel programming since the early days of Java. • Each of the four modules in the course includes an assigned mini-project that will provide you with the necessary hands-on experience to use the concepts learned in the course on your own, after the course ends. • During the course, you will have online access to the instructor and the mentors to get individualized answers to your questions posted on forums. The desired learning outcomes of this course are as follows: • Theory of parallelism: computation graphs, work, span, ideal parallelism, parallel speedup, Amdahl's Law, data races, and determinism • Task parallelism using Java’s ForkJoin framework • Functional parallelism using Java’s Future and Stream frameworks • Loop-level parallelism with extensions for barriers and iteration grouping (chunking) • Dataflow parallelism using the Phaser framework and data-driven tasks Mastery of these concepts will enable you to immediately apply them in the context of multicore Java programs, and will also provide the foundation for mastering other parallel programming systems that you may encounter in the future (e.g., C++11, OpenMP, .Net Task Parallel Library).

Visualizar o programa do curso

Habilidades que você aprenderá

Dataflow, Parallel Computing, Java Concurrency, Data Parallelism

Avaliações

4.6 (619 classificações)

5 stars
413 ratings
4 stars
152 ratings
3 stars
44 ratings
2 stars
7 ratings
1 star
3 ratings

RA

Dec 27, 2017

Very relevant concepts, explained beautifully by a very experienced professor. The assignments were good, but they could involve more programming for more practice!

QL

Jan 24, 2018

The combination of quiz and programming assignment helps understanding the course in a quite nice way. I will finish up the other two courses for sure.

Na lição

Loop Parallelism

Welcome to Module 3, and congratulations on reaching the midpoint of this course! It is well known that many applications spend a majority of their execution time in loops, so there is a strong motivation to learn how loops can be sped up through the use of parallelism, which is the focus of this module. We will start by learning how parallel counted-for loops can be conveniently expressed using forall and stream APIs in Java, and how these APIs can be used to parallelize a simple matrix multiplication program. We will also learn about the barrier construct for parallel loops, and illustrate its use with a simple iterative averaging program example. Finally, we will learn the importance of grouping/chunking parallel iterations to reduce overhead.

3.1 Parallel Loops5:51

3.2 Parallel Matrix Multiplication4:31

3.3 Barriers in Parallel Loops5:29

3.4 Parallel One-Dimensional Iterative Averaging8:07

3.5 Iteration Grouping/Chunking in Parallel Loops6:21

Ministrado por

Vivek Sarkar
Professor

Transcrição

Let us now look at a very simple example of a stencil computation,

while in practice there are many more complicated examples like this.

So the problem is as follows.

Let's say we want to solve this recurrence.

With some boundary conditions where you have,

let's say, x_sub_0 is 0,

x_sub_1 is 1 and we want to do this

for zero less than i less than n.

So i going from one to n minus one.

So we're really trying to figure out in this case an iterative averaging example

how can we compute x_i such that it's the average of its neighboring elements.

And this is in a one-dimensional array.

As I mentioned earlier,

we have stencil computations that work on much larger arrays,

multi-dimensional, that essentially follow this pattern.

So you can think about it and maybe figure

out that the solution is quite easy to work out in practice.

If we had n equal to 10,

then for i going from 0 to 10 you'd

have zero and one at the boundary and the intermediate values would be 0.1,

0.2 and so on to 0.8, 0.9.

And if you take a look, you can quickly convince yourself that each element is

indeed the average of its two neighboring elements.

Now the question is: instead of solving this in our head,

how could we write a computer program to come up with this answer?

Well, the usual approach that's taken is to have two arrays.

And this is in an algorithm that's referred to as Jacobi's relaxation.

Let's say you have an old value of the array x and you have a new value of the array x.

And then every iteration,

you're going to compute new x,

the ith element, as the average of the i minus one and the i plus 1 elements of old x.

Notice the way I've drawn it.

We can see that all these elements of new x can be computed in parallel with each other.

And when we are done, we are going to swap the pointers so that what was

earlier new x becomes old x, and what was

earlier old x can now accumulate the values for the next phase.

And the idea is if we keep repeating this computation,

we will eventually converge to the answer that we figured out analytically.

The reason the computation is important is in real world problems,

we do not have an analytical solution and we need this computation to obtain the answer.

So how would we write code to do this computation?

Well, we have an outer iteration loop,

which we can consider kind of like

a time step loop and that'll go from one to some number of times steps.

And then for each time step,

we are going to do one of these parallel steps so we can have

a for-all for i that goes from one to n

minus one and does the update with new

x -- i equal to -- and I'll just write the average of old x,

i minus one and old x, i plus one.

And when we are done with the for-all,

we have to swap the pointers.

So this is fine.

We can go through a large number of steps and each time have a for-all,

but there is an inherent inefficiency in this approach because each for-all has n

minus one tasks and that gets multiplied by the number of time steps.

So we have n steps times n minus one,

a quadratic number of tasks being created.

Now it turns out this is exactly the scenario for which barriers can help

you generate more efficient code and to do that,

what you have to do is - and it's very interesting,

you have to move the for-all loop outside.

You have the for-all loop outside and then within each iteration of the for-all,

within each parallel task i,

you do the sequential loop.

You have the sequential loop and then within the sequential loop,

you can perform this update for i,

which is the new x

sub i equals average of old x,

i minus one, and old x, i plus one.

And now here's the challenge;

you want to swap the pointers and go to

the next time step.

But before going to the next time step,

you have to make sure that all the i's have completed computing

the new x average because we are going to swap the pointers and

start overwriting what was old x, in the next phase.

And if we don't have this coordination,

then we'll end up with some data race condition.

And this, as you can guess,

is perfectly suited to a barrier operation.

So we can insert a barrier over here.

This is also referred to as a "next" operation,

which is an indication that

the for-all loop has to move to its next phase after the barrier.

And with this we ensure that all the i's

which are running in parallel for a given time step

complete the updates, wait on the barrier,

swap pointers, and then go to the next round of updates.

So what we've seen is that barriers can be used not

just to separate two phases like we saw earlier with hello and goodbye;

they can be embedded in sequential computations like for loops and while

loops and help coordinate across tasks.

Now the big advantage of doing this approach is we only

have n minus one tasks created. That's it.

We do have - the number of barrier operations is still n

minus one times n steps barrier operations,

but the number of tasks is much fewer and this is a much better tradeoff in practice

because creating tasks is much more expensive than a barrier operation,

which just involves coordination among tasks that have already been created.

So now you've seen a more interesting example using

barriers and it is indeed reflective of a lot of

real world computations and we see that barriers

can lead to more efficiency and help reduce the number of tasks that you create.

Explore nosso catálogo

Registre-se gratuitamente e obtenha recomendações, atualizações e ofertas personalizadas.

O Coursera proporciona acesso universal à melhor educação do mundo fazendo parcerias com as melhores universidades e organizações para oferecer cursos on-line.

© 2019 Coursera Inc. Todos os direitos reservados.

Baixar na App Store

Baixar no Google Play

Coursera
Sobre
Liderança
Carreiras
Lista de cursos
Certificados
Certificados MasterTrack™
Notas
For Enterprise
For Government

Comunidade
Learners
Partners
Developers
Beta Testers
Translators

Conectar
Blog
Facebook
LinkedIn
Twitter
Instagram
YouTube
Tech Blog

Mais
Termos
Privacidade
Ajuda
Acessibilidade
Imprensa
Contato
Diretório
Afiliados