A Call to Arms for Parallel Programming Standards

By Nicole Hemsoth

November 16, 2010

Although the parallel programming landscape is relatively young, it’s already easy to get lost in. Beside legacy frameworks like MPI and OpenMP, we now have NVIDIA’s CUDA, OpenCL, Cilk, Intel Threading Building Blocks, Microsoft’s parallel programming extensions for .NET, and a whole gamut of PGAS languages.

And according to Intel’s Tim Mattson, that’s not necessarily a good thing. Mattson, who is a principal engineer (and parallel programming evangelist) at the company’s Visual Applications Research Lab, says all these software frameworks are leading to what he calls “choice overload” and this concerns him greatly.

From his point of view, the road to parallel programming need to be paved with open industry standards. And today then means MPI, OpenMP, and OpenCL. Given that some of Intel’s parallel software offerings such as the Cilk Plus and Array Building Blocks are proprietary, that viewpoint sometimes puts him at odds with his own company. But Mattson’s role as an Intel researcher forces him to look beyond the one or two-year timeframe of product cycles. He’s in it for the long-term, and that means Mattson is looking at what is best for the ecosystem ten years out. “First and foremost, we have to make sure that the right standards exist and they run best on Intel products,” he says.

At SC10 this year, Mattson will be in full software evangelist mode, speaking at seven different tutorials, BoFs and panels on various parallel programming topics. Three of these are geared to fire up the troops for OpenCL, an open standard parallel programming framework for heterogeneous multicore architectures. HPCwire spoke with Mattson shortly before the conference about the importance of open standards, his unapologetic enthusiasm for OpenCL, and his open animosity for the CUDA programming language.

HPCwire: What is the significance of OpenCL and why are you devoting so much time talking about it at SC10?

Tim Mattson: I think OpenCL is perhaps the most important development in the last five, if not the last ten years. The reason I make such an over-the-top statement is that I believe the core to solving the parallel programming challenge is standards. Only an idiot software developer would write code using a propriety API. Since I don’t like to work with idiots (laughs), I want to support good software developers out there by making sure they have the full suite of standards that they need.

So we have message passing covered: MPI. It’s great. We have shared memory covered: OpenMP. It’s great. The glaring hole — because frankly I don’t think any of us saw it coming in the early 2000s — is heterogenous platforms. So we have to fill that hole, and that’s what OpenCL does. So I think it’s incredibly important because now with MPI, OpenMP and OpenCL we’ve got the space covered with these low-level basic programming standards that are required to move things forward.

HPCwire: Well as far as openness goes, NVIDIA’s CUDA programming API is available for any vendor to implement for their particular parallel hardware architecture. For example, AMD could support CUDA for their x86 and GPU platforms. So couldn’t CUDA be adopted as a standard as well?

Mattson: Well, just think about it. I can’t speak for AMD, but why would Intel put resources into an API or language that we have absolutely no say in over how it’s going to evolve? To call CUDA a standard is just insulting. It’s not a standard until the various players can all have a voice in it. It’s ridiculous. If NVIDIA was serious about it, they would create an industry working group that owns the development of CUDA’s API and languages and that has a full voice in what happens with it. Oh, by the way, that’s what OpenCL is.

HPCwire: But there is at least one example of a standard language that emerged from a vendor initiative. Java was controlled by Sun Microsystems for many years and was adopted as a commercial standard because it became so popular across the industry. Don’t you think CUDA could follow that model?

Mattson: Well I know that’s what NVIDIA would like to see happen. And yes, Java is the one instance that would call into question how absolute my statement is. Java though was coming into a very different market and was tightly associated with the Web browser — a platform that cut across the industry. And Sun showed very early on that they were willing to support it as a cross-platform language. They had Java available on x86 and Sparc and showed a willingness to work across the vendors. NVIDIA — rationally, by the way — isn’t doing this with CUDA.

When you look closely at OpenCL, it covers everything CUDA can do and more. OpenCL has all the key vendors and covers a much wider space than CUDA. We’ve got the embedded people, the cell phone vendors, and game vendors all involved. So OpenCL is the right way to go; CUDA is the wrong way to go.

HPCwire: In the high performance computing community, though, there has been criticism that OpenCL doesn’t deliver the kind of performance required for HPC codes. Do you think that’s a fair assessment?

Mattson: That’s a statement that’s both true and false. There’s nothing pathological in the definition of OpenCL that prevents it from being every bit as efficient as CUDA. The thing about OpenCL is that it’s young; it just hasn’t been out very long. So it really comes down to the vendors as far as the quality of their implementations.

I think it’s important for the programmers out there — and let’s face it, they are the end user community for these technologies — to steer things in the right direction by insisting on standards. Look at how MPI and OpenMP came into existence. In both those cases, the user community insisted that these standards be the foundation of the software ecosystem, and the vendors stood behind them. We need people to do the same thing here and not get caught up with point solutions.

If NVIDIA engineers spent as much time optimizing OpenCL, it would run as fast as CUDA. So the performance arguments don’t hold a lot of sway with me, except when someone can say this feature of the language as defined is fundamentally going to be inefficient regardless of the quality of implementation. When people find those, we in the OpenCL group take it very seriously.

We’re roughly on a two-year cadence of coming out with new releases of the OpenCL spec, and we’re very focused on finding the weakness in OpenCL and aggressively evolving the language and to stay right in line with the latest hardware trends.

HPCwire: There are plenty of other languages that address multicore parallelism, some of which have been introduced by Intel. How does OpenCL fit in?

Mattson: Let me be really clear. There are three distinct standards that address multicore. MPI, for example, works great on multicore. OpenMP, if you have a shared address space, works really well too. And OpenCL covers heterogenous architectures. It’s really the trio that I’m pushing and Intel is 100 percent behind them.

On the other hand, yes, there is a trend that I find deeply disturbing of vendors wanting to distinguish themselves by creating new languages and proprietary APIs. It’s disturbing because time spent on a new language or proprietary API is time not spent on improving and establishing these standards. So this is where I’m kind of at odds with some of my colleagues at Intel. That’s just the way it goes.

Let’s face it. Vendors, left to their devices, want people to adopt a proprietary API that lock them into their platform. That’s not bad. NVIDIA is completely rational in wanting to lock people into their platform with CUDA. If I was working at NVIDIA, I’d probably be trying to do the same think. I think it’s up to the user community to refuse to let users get away with that game. They can do that by insisting on standards or open source solutions.

The three standards I mentioned are where I think most of the resource should go. But Intel did release Task Building Blocks –TBB — as open source. That was a very responsible thing to do. I was very excited, as was the TBB team, when that happened.

HPCwire: Another language is Ct, which started as a research project at Intel and has now been commercialized. How does that fit in to this parallel language ecosystem?

Mattson: Ct, which by the way, is now called Array Building Blocks, is a higher level abstraction of parallelism. While I’m a huge supporter of what they are doing with Array Building Blocks, [as for] how useful it will be in the marketplace, I’m not sure because it is proprietary. But I think some of the optimizations it does under the covers is very important. There are a lot of really important things about that project.

But I think we should distinguish creating good technologies versus confusing the market by having too many options out there. In 2004, if you wanted to do parallel programming you had Windows threads, pthreads on Linux, OpenMP, and MPI. Five options was fine. Now there are a dozen or more parallel programming languages out there. So I think we’re losing ground. I think choice overload is real. And that concerns me deeply.

HPCwire: Do you think parallelizing established languages like Java and Python is a positive development?

Mattson: Let me tell you where I think things are going and where we’ll be in 10 years. The question is do we get there cleanly or do we get there with messy detours along the way. Ultimately I think we have to raise the level of abstraction, which is what you see with these efforts around building parallelism into Python. We need to focus on the higher level frameworks that people are increasingly using to write software.

This is really what I spend the bulk of my time doing in my personal research, and with a group at UC Berkeley — to define patterned languages from which we can derive the frameworks, which then map down to the lower-level languages. I really want to make it so that only a small number of performance-oriented, efficiency layer programmers worry about these low-level languages — OpenCL, MPI, OpenMP, or TBB. But beyond that, people need to have some higher level framework they can work with. A parallel Python project like Copperhead is such an example. I’m very excited about it because I think that’s clearly the direction things are moving.

I learned this most clearly looking at the gaming industry, because that industry has been the leader in adopting multicore, and I mean adopting multicore as a successful business venture. Researchers have been playing with it for a long time, but in terms of creating multicore software, selling it, and building a profitable businesses around it, the gaming industry has led the charge.

They have these separations of concerns very sharply defined, and it works extremely well for them. Most of their programmers work in a higher level scripting language or with collections of libraries written in C++. And then they have a small number of “technology programmers” which are on the order of 10 percent of their software developers. They’re the guys who do the low-level stuff. And I think that kind of separation of concerns is what’s absolutely critical.

HPCwire: So you think higher level frameworks will be key to enabling these low-level parallel programming APIs you’re talking about?

Mattson: When we were sitting around creating OpenCL, we explicitly talked about that as our goal. In fact, you’ll find some places in the spec where we describe OpenCL as a hardware abstraction layer. We’re perfectly aware that OpenCL is obnoxiously low-level. It exposes so many details of the underlying platform. We achieve extreme portability by exposing everything and abstracting as little as we can. The reason we think that’s the right thing to do is because we view OpenCL ultimately as being a target for higher level frameworks. It’s young, so those higher level frameworks don’t exist yet, but I think they will and I think that will be the long range trend, not just for OpenCL, but for all these parallel languages.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing power it brings to artificial intelligence.  Nvidia's DGX Read more…

Call for Participation in Workshop on Potential NSF CISE Quantum Initiative

March 26, 2024

Editor’s Note: Next month there will be a workshop to discuss what a quantum initiative led by NSF’s Computer, Information Science and Engineering (CISE) directorate could entail. The details are posted below in a Ca Read more…

Waseda U. Researchers Reports New Quantum Algorithm for Speeding Optimization

March 25, 2024

Optimization problems cover a wide range of applications and are often cited as good candidates for quantum computing. However, the execution time for constrained combinatorial optimization applications on quantum device Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at the network layer threatens to make bigger and brawnier pro Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HBM3E memory as well as the the ability to train 1 trillion pa Read more…

Nvidia Appoints Andy Grant as EMEA Director of Supercomputing, Higher Education, and AI

March 22, 2024

Nvidia recently appointed Andy Grant as Director, Supercomputing, Higher Education, and AI for Europe, the Middle East, and Africa (EMEA). With over 25 years of high-performance computing (HPC) experience, Grant brings a Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HB Read more…

Nvidia Looks to Accelerate GenAI Adoption with NIM

March 19, 2024

Today at the GPU Technology Conference, Nvidia launched a new offering aimed at helping customers quickly deploy their generative AI applications in a secure, s Read more…

The Generative AI Future Is Now, Nvidia’s Huang Says

March 19, 2024

We are in the early days of a transformative shift in how business gets done thanks to the advent of generative AI, according to Nvidia CEO and cofounder Jensen Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Nvidia Showcases Quantum Cloud, Expanding Quantum Portfolio at GTC24

March 18, 2024

Nvidia’s barrage of quantum news at GTC24 this week includes new products, signature collaborations, and a new Nvidia Quantum Cloud for quantum developers. Wh Read more…

Houston We Have a Solution: Addressing the HPC and Tech Talent Gap

March 15, 2024

Generations of Houstonian teachers, counselors, and parents have either worked in the aerospace industry or know people who do - the prospect of entering the fi Read more…

Alibaba Shuts Down its Quantum Computing Effort

November 30, 2023

In case you missed it, China’s e-commerce giant Alibaba has shut down its quantum computing research effort. It’s not entirely clear what drove the change. Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

DoD Takes a Long View of Quantum Computing

December 19, 2023

Given the large sums tied to expensive weapon systems – think $100-million-plus per F-35 fighter – it’s easy to forget the U.S. Department of Defense is a Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Intel’s Server and PC Chip Development Will Blur After 2025

January 15, 2024

Intel's dealing with much more than chip rivals breathing down its neck; it is simultaneously integrating a bevy of new technologies such as chiplets, artificia Read more…

Baidu Exits Quantum, Closely Following Alibaba’s Earlier Move

January 5, 2024

Reuters reported this week that Baidu, China’s giant e-commerce and services provider, is exiting the quantum computing development arena. Reuters reported � Read more…

Leading Solution Providers

Contributors

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock 1179408610

Google Addresses the Mysteries of Its Hypercomputer 

December 28, 2023

When Google launched its Hypercomputer earlier this month (December 2023), the first reaction was, "Say what?" It turns out that the Hypercomputer is Google's t Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

Google Introduces ‘Hypercomputer’ to Its AI Infrastructure

December 11, 2023

Google ran out of monikers to describe its new AI system released on December 7. Supercomputer perhaps wasn't an apt description, so it settled on Hypercomputer Read more…

China Is All In on a RISC-V Future

January 8, 2024

The state of RISC-V in China was discussed in a recent report released by the Jamestown Foundation, a Washington, D.C.-based think tank. The report, entitled "E Read more…

Intel Won’t Have a Xeon Max Chip with New Emerald Rapids CPU

December 14, 2023

As expected, Intel officially announced its 5th generation Xeon server chips codenamed Emerald Rapids at an event in New York City, where the focus was really o Read more…

IBM Quantum Summit: Two New QPUs, Upgraded Qiskit, 10-year Roadmap and More

December 4, 2023

IBM kicks off its annual Quantum Summit today and will announce a broad range of advances including its much-anticipated 1121-qubit Condor QPU, a smaller 133-qu Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire