The Future of User-Directed SMP Parallel Programming

By Michael Wolfe

September 1, 2006

OpenMP (http://www.openmp.org) is an open standard for C, C++, and Fortran pragmas and directives to support portable shared-memory parallel programming. The current standard is OpenMP 2.5, which was defined by a language committee comprising users and vendors, and is available on many current high performance platforms.

The OpenMP language committee is now discussing many issues regarding features to include in the next version of the standard, OpenMP 3.0. This article summarizes the proposed extensions likely to be incorporated in 3.0, more extensive issues we are working to define for inclusion in 3.0, and other issues that for one reason or another are not being discussed or were explicitly deferred until some future revision of the OpenMP standard. The general goal of the OpenMP language committee is to complete the major work for the OpenMP 3.0 standard by Supercomputing 2006 in November, and prepare it for public comment shortly thereafter.

There are several straightforward issues that will improve both the ease-of-use and efficiency of OpenMP. There is broad agreement on these features, so it's really just a matter of defining them carefully and working them into the standard. Examples include:

Loop Collapse: the OpenMP 3.0 standard will likely include a way to generate parallel code for nested loops as efficiently as if they were a single loop. The proposal will probably look something like:

        #pragma omp parallel do collapse(2) shared(a,b,c) private(i,j)
        for( i = 0; i < n; ++i ){
            for( j = 0; j < m; ++j ){
               a[i][j] = b[i][j] + c[i][j];
            }
        }

with the restriction that the loops are tightly (perfectly) nested.  The iterations of the nested collapsed loops will be scheduled as one block of n*m (in this example) iterations among the parallel threads. An alternate proposal named the loop indices instead of giving the number of loops to collapse, but this failed because it would seem to allow parallelization of nonadjacent loops.

Stack Size Control: When OpenMP parallel threads are fired up, the stack for each thread is allocated from the heap. Each implementation has a different default size for this stack. If one of these stacks overflows, such as with deeply nested routines, the application will fail. The OpenMP 3.0 standard will likely include a way for the user to set the stack size, probably using an environment variable:

        export OMP_STACK_SIZE 64M

and to query the stack size with an API routine.

Idle Thread Control: Between parallel regions, the threads created for the parallel region are usually parked somewhere. Some implementations use a busy-wait scheme, so the latency of starting the next parallel region is as low as possible. Other implementations use a sleep facility, so the threads don't interfere with other work going on in the system. We will give the user some control over the idle behavior. One issue under discussion is whether to distinguish between fine grain waiting behavior, such as at a synchronization point within a parallel region, and coarse grain waiting, such as between parallel regions. The OpenMP 3.0 standard will likely have an environment variable, OMP_WAIT_POLICY, that a user can set with values such as BUSYWAIT, SLEEP, YIELD, to get the desired behavior, though the actual wording is yet to be defined.

Multiple Internal Control Variables: The OpenMP language standard defines several “internal control variables,” settings within the implementation, which can be set and/or queried by a program or user. These include parameters like how many threads to use in a parallel region, whether nested parallelism is enabled, the default loop scheduling policy, and so on; the idle thread policy discussed above would be added to this list. The standard doesn't define what happens when a program tries to reset the value of these variables within a parallel region. This might occur, for instance, when there is a call from within a parallel region to a subroutine library that has its own parallel regions. The OpenMP 3.0 standard will likely define that each thread gets its own copy of the internal control variables, and that each team created at a parallel region inherits the values from the master thread; querying and resetting the value of any of these will be well-defined wherever it occurs.

Private Variable Master Copy: When entering a parallel region with a private variable, the standard allows for the implementation to reuse the original variable storage for the master thread's private copy. This was a performance issue, but causes no end of misunderstandings for users. The OpenMP 3.0 standard will likely disallow the storage reuse, for many reasons, not the least of which is the behavior of C++ constructors and destructors, as well as the potential for memory leaks with C pointers and Fortran allocatables.

Unsigned int: OpenMP 2.5 allows 'int' types for parallel loop index variables. C and C++ also have 'unsigned' types, as well as 'long' and sometimes 'long long'. The OpenMP 3.0 standard will likely allow any supported integer datatype as a loop index.

Static schedule: The simplest loop schedule is the static schedule, defined such that the iterations are divided into chunks of approximately equal size, and each thread gets at most one chunk. Given four threads, the schedule for the following loop

        for( i = 0; i < 10; ++i ){
            dowork(i);
        }

could be 3/3/3/1 (three threads get 3 iterations, the last thread gets 1 iteration), 3/3/2/2, 4/4/2/0, or even 1/4/4/1. There was some discussion that allowing an implementation more freedom would let it align the chunks to cache line boundaries, avoiding false sharing between processors and improving performance, even though unbalanced. However, the OpenMP user community requested a more strict definition, to allow users to place the NOWAIT directive between loops, allowing early finishing threads to continue to the next loop. The static schedule will likely be defined such that only the first schedule listed above will be compliant in OpenMP 3.0.

In addition to these straightforward issues, there are a number of nontrivial issues the OpenMP language committee is working to address. Some of these are potentially very important to the future success of OpenMP, but they are difficult to define and in some cases may conflict with certain vendor-defined extensions to OpenMP. These are issues that the OpenMP language committee would like to include in OpenMP 3.0, but which require extensive work and discussion, which is ongoing. Examples include:

Task Queues: The Intel compiler has for some years had a task queue facility, whereby a program can define a block of work that can be enqueued on a queue of tasks and executed in parallel with other tasks. This is quite good for parallelism across linked list members, for example, or for a recursive traversal of a data structure. One proposal looks much like the following:

        #pragma omp taskqueue private(ptr) shared(head)
        {
            for( ptr = head; ptr; ptr = ptr->next ){
                #pragma omp task capture(ptr)
                { handle(ptr); }
            }
        }

There is quite a lot of active discussion about the details of this issue, such as when a thread that starts a task can enqueue more tasks, whether a thread that starts a task can pause and start a different task, whether a task that was paused can be picked up and continued by a different thread, whether a task that has created threads needs to wait at the end of the routine in case the subtasks refer to any local state of that routine. The real problem is the thread-centric execution model of OpenMP and a mismatch with a tasking model. This is possibly the biggest change that may be included in OpenMP 3.0, and arguably the most generally useful.

Constructors and Destructors: The standard is very lax in defining when C++ constructors and destructors are invoked for private variables. The OpenMP 3.0 standard may include a much more detailed definition.

Memory Model: There is a memory model in the OpenMP 2.5 specification, but it really needs to be carefully revised. There is general agreement that this should be included in the OpenMP 3.0 specification, but time constraints on the proposed release schedule for the standard may cause this to slip to a future revision.

Schedule Kinds: Loop schedules in OpenMP 2.5 are static (each thread gets fixed chunks of the same size), dynamic (also called self-scheduling in the literature), guided (guided self-scheduling in the literature), and runtime (let the user decide at runtime using an environment variable). As mentioned above, OpenMP 3.0 will likely include tightening up the definition of the static schedule. In addition, some new loop schedule types are under discussion:

  • Affinity scheduling, which the IBM compiler implements, and which prefers to keep similar chunks on the same threads across loops, hoping for cache locality.
  • Auto scheduling, which uses a runtime mechanism to re-balance the load between threads across multiple executions of the same loop.
  • Named schedules, which allow a user to reuse a schedule across multiple loops, again to control load balance or cache locality.
  • General block schedule, which allows a user to define how many iterations are allocated to each thread, as flexible as possible.

Fortran Allocatable, Assumed Shape: OpenMP 2.5 does not allow Fortran allocatables or assumed shape arrays in all places where it might make sense. This may be addressed in the OpenMP 3.0 standard.

Conclusion

Clearly there are a large number of issues the OpenMP language committee is actively working on for OpenMP 3.0 and many additional issues they are considering for future versions. OpenMP has been quite successful as a de facto SMP parallel programming standard since the first revision of the standard was released in 1997. Almost every major compiler supplier supports OpenMP, and a large number of ISV and proprietary applications have been parallelized for shared-memory systems using OpenMP, including many mainstream technical applications in MCAE, CFD, computational chemistry, and geophysical processing.

A significant part of the appeal of OpenMP is its incremental nature, which allows developers to parallelize small portions of an application. However, OpenMP is also designed to enable entire applications to be parallelized in a comprehensive way. For this reason, I expect OpenMP will continue to be appealing to an ever broadening set of developers looking to parallelize applications for today's multi-core processors.

—–

Dr. Wolfe has 30 years professional compiler experience in both academia and industry, and has been with The Portland Group (PGI) for the past ten years. He has many publications, including two books, on high performance and parallel compiler optimizations and serves as PGI's representative on the OpenMP Architecture Review Board.

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!

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion XL — were added to the benchmark suite as MLPerf continues 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 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…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion 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…

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