Why is Parallel Computing Hard?

There are plenty of issues with parallel programming. Breaking up the problem is often the most important and complex step, especially when the parallelism is not obvious. As we are rooted in a world of sequential programming, conceptualizing the parallelization of tasks that lend themselves to sequential programming is tough. This can require not only the reworking of code, but redesigning the entire process of solving a problem.

Even in problems that lend themselves to parallelism, exploiting the parallelism can be tough. Even if you know the best and fastest algorithm for solving a data parallel problem, it isn't always possible to translate that to an efficient program. For instance, if I want to multiply two matrices with 100k x 100k dimentions, I can't just spawn all the threads I would need. If I were using POSIX threads to calculate one cell of the result matrix each, I would spend more time creating threads and allocating resources than actually doing the computation. I've got to take the resources I have and use them to the best of my ability. Though I can do matrix multiplication in parallel, I have to be careful about how I break up the problem and I can't exploit all the parallelism possible because of the tools I normally work with.

We are also limited in terms of hardware resources. With only a few processors available for general purpose programming, even if the software overhead weren't an issue we couldn't actually get any speed up from parallelizing beyond a certain point. This not only means that we can't exploit tons of parallelism even if the algorithm lends itself to it and this discourages programmers from thinking in terms of parallelism.

How Does OpenCL Help?

What if we had not only a pool of hardware resources hundreds wide that could handle thousands of threads in flight at a time with no software overhead? Well, we do: it's called a GPU. And if we could use the GPU for processing, then we could spawn a bunch of threads and really chew through the matrix multiplication we talked about earlier (or whatever). We might still have to be concerned about how many hardware resources we have in order to best map the problem to the specific device in the system. And we still have the problem of actually spawning, managing and running threads on the GPU hardware.

But what if we could write a special function, called a kernel, that can instantly be spawned hundreds or thousands or millions of times and run on different data all without needing to handle creating and managing all the threads ourselves. And what if we didn't need to worry about how to break up our problem and left actually determining how to handle allocating threads to the runtime? Well, now we have a solution: that's OpenCL.

The GPU is the vehicle for exploiting data parallelism. But before now our vehicle has run like a train on a track called real-time 3D graphics acceleration. OpenCL removes the track and the limitations and builds in a steering wheel developers can use to take the GPU (and other parallel devices) anywhere a programmer can imagine.

Index Open, Closed, Proprietary ... Sorting out the Confusion


View All Comments

  • v12v12 - Wednesday, January 07, 2009 - link

    Testing123, ignore plz Reply
  • corporategoon - Tuesday, January 06, 2009 - link

    Did this article go through an editor? Reply
  • chizow - Friday, January 02, 2009 - link

    Kind of surprising you didn't directly address this given the amount of FUD being thrown around with regards to PhysX, particularly from AMD and its supporters. You indirectly answered what I had already suspected however, that given Nvidia has stated they plan CUDA to be fully portable to both OpenCL and DX11 there should also be no portability issues for AMD and Brook+:


    AMD could make an investment in the CUDA for C language and create either their own compiler (nothing is stopping them). But then you still have the same problem of interoperability as if NVIDIA implemented Brook+. If NVIDIA or AMD want to make their solution work with the other guy, they would need to write a wrapper to translate CAL to PTX or PTX to CAL.

    I'm guessing the unfinished thought from the first sentence should read something like "or write a CUDA to Brook+ wrapper" as thats essentially what the last part suggests. Since both vendors will write wrappers for their code to OpenCL, perhaps this wrapper could pull double duty, although it would double the amount of transcoding needed. Less than efficient for sure, but certainly better than a complete impasse due to incompatibility.

  • ltcommanderdata - Friday, January 02, 2009 - link

    Are you suggesting that hardware PhysX acceleration will come to AMD GPUs as soon as nVidia and AMD enable hardware OpenCL support? Because I don't think it's that simple.

    nVidia seems to have rebranded the meaning of CUDA. Maybe it's all just marketing speak, but CUDA before seemed to mean using nVidia GPUs for GPGPUs operation in general. But now since OpenCL, CUDA seems to more specifically related to the GPGPU interface to nVidia GPUs with languages being separate on top, namely OpenCL, DX11 and C for CUDA. If PhysX is written in C for CUDA, which it no doubt is seeing there wasn't anything else available up to now, then adding support for the OpenCL language in the CUDA interface layer won't help get PhysX supported on AMD GPUs. PhysX will still be written in nVidia's proprietary language which AMD GPUs can't understand. To support AMD GPUs, either nVidia will have to rewrite PhysX from C for CUDA to OpenCL, which would be awfully generous of them or AMD will have to make a C for CUDA to CAL translator and hope PhysX doesn't have any nVidia hardware specific optimizations, which it no doubt has, to mess things up.
  • apanloco - Friday, January 02, 2009 - link

    Anyone knows if multiple applications can take advantage of OpenCL at the same time? I think OpenGL is exclusive to one application, but if OpenCL is used by regular applications this could be a problem?
  • yyrkoon - Thursday, January 01, 2009 - link

    "With R580 AMD (then ATI) actually published part of their ISA and called the initiative CTM (for Close to Metal). Before we had a beta version of CUDA, we had folding@home GPU accelerated on R520 and R580"

    I also read an interview through gamedev.net where ATI was emulating Direct 3D 10 calls in hardware on one of their x1900xtx's ( Direct 3D 9 hardware )long before I heard about folding@home on the GPU. I remember being so impressed with the technology, that I could not wait until Vista + Directx 10 titles became available. Too bad that there are so few ( if any ) titles that currently take advantage of this technology in the ways I had hoped. Hopefully that will change soon.
  • ltcommanderdata - Thursday, January 01, 2009 - link


    It's interesting that you mentioned that AMD and nVidia look to be continuing to push their proprietary GPGPU solutions, but AMD has actually made statements they are abandoning their proprietary CTM GPGPU implementation and are moving fully to OpenCL. Admittedly, its probably just a realization that CTM isn't taking off as fast as CUDA and it's in their best interest to push OpenCL. In comparison, nVidia will continue to develop their own CUDA implementation alongside OpenCL.

    I wonder if you can get a statement from nVidia whether they will move PhysX to OpenCL? Right now I believe PhysX is written in C for CUDA and of course requires nVidia GPUs for hardware acceleration. If they moved to OpenCL, then AMD GPUs would support it as well. Although perhaps nVidia prefers to keep PhysX to themselves as a product differentiator.

    It'd also be interesting if you could ask AMD whether older GPUs like the X1600, X1800, and X1900 will be supported in OpenCL? You already pointed out in your article that the RV530, R520, and R580 had GPGPU folding@home clients so they are certainly capable of GPGPU operation. It'd probably be in ATI's own interest to have as large an OpenCL base as possible and ATI's original FireStream dedicated GPGPU card was R580 based as well. Apple could probably help them as well seeing the number of X1600 and X1900 used in various iMac, MacBook Pro, and Mac Pro generations that could use support for OpenCL in Snow Leopard.

    And I agree with melgross that it's strange Apple got no mention in the article seeing that they pretty much developed OpenCL, then submitted it to Khronos, and was no doubt a major driving force behind the quick ratification in order to get it ready for Snow Leopard. And I believe Apple's Aaftab Munshi was the chair of the OpenCL working group.
  • danger22 - Thursday, January 01, 2009 - link

    i am looking forward to the day when I can run my finite element simulations on my GPU. come on Ansys its time for a GPGPU Multiphysics! Reply
  • Amiga500 - Thursday, January 01, 2009 - link

    Same boat, same boat... with both CFD and FEA.

    Have you heard of FEAST-GPU (from Dortmund university)?

    Its a GPU accelerated FE package - unfortunately it isn't out in the public domain yet.

    Anyhow - from my own digging, I'm not sure if the CPU is a major bottleneck for FE simulations - a lot of what I see tends to point towards the hard-drive and I/O performance.
  • Sheep100 - Sunday, January 04, 2009 - link

    If you provide enough RAM to the analysis you definitely end up CPU limited for single core runs. We have 24 - 32 GB per node for Abaqus and Nastran analyses. The nodes get RAM - bandwidth limited when stepping up the number of cores used or the number of concurrent runs on a node. We are looking forward to the core i7/Nehalem Xeon systems coming soon that will provide a big improvement here. (These codes run slower on Opteron cores.)

    GPGPU versions of Abaqus, Nastran & Ansys would be very interesting given the large memory bandwidth available on the high end cards. I suspect that re-writing & validating the various solver algorithms to target OpenCL would be a long process. I'm also unsure how possible it is to get data parallelism out of them since the scaling rate of Abaqus, for example, on multi-core systems, even with good bandwidth, is not anywhere near linear. Although this might just highlight the deficiency of the current method of extracting parallelism.

Log in

Don't have an account? Sign up now