parallel computing in C++

[carlosp]

I am programing something that could greatly benefit from parallel
computing in C++, in parts of the program, where user defined classes
are passed. I have googled but I cannot find the proper tool. Any
experienced programmer could give me a hint, namely some good
keywords?


Thanks!

 

[Ian Collins]

I am programing something that could greatly benefit from parallel
computing in C++, in parts of the program, where user defined classes
are passed. I have googled but I cannot find the proper tool. Any
experienced programmer could give me a hint, namely some good
keywords?

You don't have a ask twice!

Look for openMP.

 

[Juha Nieminen]

Look for openMP.

This is a good tutorial for those fluent in C++:

http://bisqwit.iki.fi/story/howto/openmp/

 

[Rui Maciel]

I am programing something that could greatly benefit from parallel
computing in C++, in parts of the program, where user defined classes
are passed. I have googled but I cannot find the proper tool. Any
experienced programmer could give me a hint, namely some good
keywords?

What's wrong with pthreads?


Rui Maciel

 

[Rui Maciel]

It's very low-level, and does not provide most of the tools one need for
real-world parallel processing: thread pools

Isn't that a common design pattern that is easy to implement?

data structures that
support concurrent access

What's wrong with mutexes and read-write locks?



Rui Maciel

 

[Chris M. Thomasson]

Isn't that a common design pattern that is easy to implement?

Of course.

What's wrong with mutexes and read-write locks?

Nothing. However, if your looking for extremely scaleable high-performance
data-structures, well... That's another story... Perhaps best discussed over
on `comp.programming.threads'.

 

[SaticCaster]

OpenMP Parallel Programming
http://www.viva64.com/links/parallel-programming/

 

[Chris M. Thomasson]

Not really.

You can create full-blown lock-based work-stealing schedulers in POSIX that
have pretty darn good scalability. Wait-free work-stealing algorithms are
another story... Here is a pretty good one:

http://groups.google.com/group/comp.programming.threads/browse_frm/thread/8ad297f61b369a41

In particular, assigning tasks to threads at random kills cache coherency
and performance.

That depends on several factors. Anyway, a simple way to get good cache
performance is to batch up similar actions together and execute them all on
a single CPU:

http://www.usenix.org/events/usenix02/full_papers/larus/larus.pdf

You can write "a thread pool," but it takes multiple person-months to get
something competitive with COTS solutions.

Sometimes. Intel TBB and Clik++ are both based on work-stealing, and those
algorihtms are not rocket science, IMVHO of course...

 

[Chris M. Thomasson]

Sometimes. Intel TBB and Clik++ are both based on work-stealing, and those
algorihtms are not rocket science, IMVHO of course...

However, COTS are OH SO CONVENIENT. AFAICT, Intel TBB is a very good choice.

:^)