10

Providers of new computer architectures regularly try to introduce new programming models, e.g. recently CUDA/OpenCL for GPGPUs, and displace C/POSIX as the control interface to platform parallelism. (Poss & Koening, AM3: Towards a hardware Unix accelerator for many-cores, 2015)

Why do architecture designers try to design new programming models to displace C/POSIX for parallel computing? Is C/POSIX not well suited to multiprocessors or did the original authors of C/POSIX not have the need for parallel computing in mind at C/POSIX design time? Or is it the case that programmers need more capabilities than C/POSIX can deliver,thus resorting to new designs e.g. CUDA/OpenCL, etc.?

  • 7
    Remember, software is an abstraction of hardware. If hardware changes too much, then the software abstraction might not be a good one anymore. I believe this is certainly true when considering using POSIX threads for a GPU, but I will leave it to someone else to explain more fully in an answer. – user22815 Dec 2 '15 at 20:22
10

Compare POSIX threads and Grand Central Dispatch, for example. I have code that dispatches to four threads in eight lines of code. With POSIX, that would all be an absolute nightmare.

On the other hand, CUDA / OpenCL are not about multithreading at all, but about using massive vector abilities. (They can do multithreading as well, but vectorizing is the important thing).

9

There is a distinction between SIMD parallel programming and the more traditional parallel programming model that POSIX uses.

SIMD is the model that CUDA, OpenCL, etc. use. There is a single set of instructions that are executed simultaneously by many threads, each operating on their own pool of data. This is very useful for thing like 3-D graphics, where the same transformations are applied to a large number of points.

The POSIX model assumes that each thread runs asynchronously, and each thread can potentially execute totally different code.

Both models have their strengths and weaknesses -- that's why they are different. POSIX is much more flexible, but CUDA/OpenCL/etc. can take advantage of specialized hardware, running thousands of (usually simpler) threads at a time.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.