Compilers and More: Parallel Programming Made Easy?

By Michael Wolfe

September 2, 2008

Look at all the current research projects aimed at exactly that: making parallel programming easy!

The NESL language aims to “make parallel programming easy and portable.” More precisely, its goal is to allow programmers “to write parallel code as concisely and clearly as sequential code, while achieving close to peak performance.”

Microsoft’s Parallel Computing Platform team aims to provide a runtime that provides support for parallelism, models, libraries and tools that “make it easy for developers to construct correct, efficient, maintainable and scalable parallel programs.”

The CxC (C by C) language has one-sided communication operations, “which makes parallel programming easy and efficient.”

My alma mater, the University of Illinois, has a new Universal Parallel Computing Research Center, whose goal is simply “Making parallel programming easy.” Marc Snir, the Computer Science Department head, said the goal for this effort is to “make ‘parallel programming’ synonymous with ‘programming’.”

The Parallel Computing Lab at the University of California at Berkeley has adopted the goal to “make it easy to write correct programs that run efficiently on manycore systems, and which scale as the number of cores doubles every two years.”

Intel’s Threaded Building Blocks (TBB) extends C++ for parallelism in an easy to use and efficient manner.

Tim Mattson (Intel) points out that in “our quest to find that perfect language to make parallel programming easy,” we have come up with an alarming array of parallel programming choices: MPI, OpenMP, Ct, HPF, TBB, Erlang, Shmem, Portals, ZPL, BSP, CHARM++, Cilk, Co-array Fortran, PVM, Pthreads, Windows threads, Tstreams, GA, Java, UPC, Titanium, Parlog, NESL, Split-C, and on and on.

Tim and I are colleagues on the OpenMP Language committee, which recently finalized the OpenMP 3.0 standard. OpenMP’s mission is to define a “portable, scalable model that gives shared-memory parallel programmers a simple and flexible interface for developing parallel applications for platforms ranging from the desktop to the supercomputer.” Is simple even easier than easy?

Every time I see someone claiming they’ve come up with a method to make parallel programming easy, I can’t take them seriously. First, “making parallel programming easy” must be harder than “making programming easy,” and I don’t think we’ve reached that first milestone yet. Yes, I can (and do) knock off quick programs to search or compute something simple, but then I can knock off a quick project to build a bridge over the seasonal stream in our backyard, too. I wouldn’t trust that bridge to freeway traffic, and I wouldn’t trust that quick program to solve anyone else’s problem, either.

All this is folly. I agree with Andrew Tanenbaum, quoted at the June 2008 Usenix conference: “Sequential programming is really hard, and parallel programming is a step beyond that.” Yes, programming is really hard. Look at the programming effort it takes to produce a well-supported world-class application. If it were easy, Microsoft wouldn’t need legions of programmers to develop and support its software. If sequential programming were a solved problem, we wouldn’t have the relatively recent introductions of new and successful languages, such as Java, and C#.

Consider any software development project. Typically it begins with discussion of interfaces (with other software, with data, with users), standards (programming language, operating system, formats), and requirements (functionality and performance). Often, a prototype implementation is developed and tested. One of the tests will involve performance, especially at the limits of the expected input. If the performance is satisfactory, no further tuning is necessary. If not, the team will look at alternative implementations, different algorithms, data structures, shortcuts, etc.; one of the schemes will be to use more parallelism.

In fact, the ONLY reason to consider parallelism is for better performance. Parallelism by itself doesn’t deliver any new features or functionality; it may allow you to deliver new functionality because of improved performance, but the parallelism itself didn’t create the functionality. Most of the programs I write don’t need parallelism because they don’t take long enough to matter. Only those with large datasets or lots of computation are even candidates. I’d be wasting my time and my employer’s resources adding complexity to my program in order to use parallelism that I don’t need.

Moreover, parallelism does not equate to performance. The focus needs to be not on parallelism, but on performance, where parallelism is one of the tools to get it.

Time and again we hear advocates claiming that if you just use their compiler, language, library, tool, methodology, etc., they will guarantee good performance now and forevermore. Yet each time around, the methods are limited to the technology of the day. Let’s look at an example algorithm, matrix multiplication:

   for i in 1:n
     for j in 1:n
       for k in 1:n
         c(i,j) = c(i,j) + a(i,k)*b(k,j)

Early compilers were tuned to optimize just such programs. As pipelined functional units were introduced, libraries were written to replace the inner kernel loops, such as STACKLIB and the BLAS; we could write this loop using a DAXPY call:

   for j in 1:n
     for k in 1:n
       daxpy( n, b(k,j), a(1,k), 1, c(1,j), 1 )
       ! equivalent to:
       ! for i in 1:n
       !   c(i,j) = c(i,j) + b(k,j)*a(i,k)

The library routines were rewritten in vector mode in the 1970s. However, it was soon realized that you could achieve even higher vector performance, called supervector performance, if you optimized the inner two loops to take advantage of vector register locality. Thus came the level-2 BLAS, implementing matrix-vector operations. Matrix multiplication turned into a loop around a DGEMV call:

   for j in 1:n
     dgemv( ‘n’, n, n, 1.0, a(1,1), n, b(1,j), n, 1.0, c(1,j), 1 )
     ! equivalent to:
     ! for k in 1:n
     !   for i in 1:n
     !     c(i,j) = 1.0*c(i,j) + 1.0*b(k,j)*a(i,k)

This was sufficient until the microprocessor revolution, where cache behavior dominated the processor performance. We then were given the level-3 BLAS, implementing matrix multiplication in a single call to DGEMM, which could then be appropriately tiled, unrolled, vectorized, and optimized for each machine. Now we’re in the multicore or manycore era, and who knows where it will lead in the next decade or two? Given how well we’ve predicted what the machines of the future will look like, can we design the universal programming model today?

One aspect of computer science is algorithm analysis, studying the computational and memory requirements of an algorithm. We learn how to reverse a linked list in linear time and constant space. We learn about the differences between bubble sort (O(N^3)), quicksort (O(N log N) on average, easily parallelizable) and heapsort (O(N log N) worst case, and my personal favorite). We learn to pay attention to the hidden constants in the big-O notation, since the constant factor can dominate the computation cost for reasonable inputs; thus, O(N^3) is almost always worse than O(N^2), but O(N log N) with a small constant may be better than O(N) with a big constant, for the expected cases.

For instance, in compilers, no one ever uses the linear time algorithm for computing dominators, because the constant factor is too big. We learn that the analysis usually translates directly to the actual performance. And we learn tricks, like hash table lookup, to work around difficult performance problems. None of this depends on the particular programming language or processor; the von Neumann model has served us well.

Algorithm analysis for parallel computing studies parallelism models, such as Bulk Synchronous Parallel (BSP), Communicating Sequential Processes (CSP), Data-Flow, and so on. These models make assumptions about the cost (or lack thereof) for synchronization and communication. Until we have machines that implement these models more closely, we need to take into account the cost of the virtualization as well. It doesn’t matter how efficient an algorithm is in some abstract machine model if the implementation of the abstract model is itself expensive.

The current “parallelism crisis” can only be resolved by three things. First, we need to develop and, more importantly, teach a range of parallel algorithms. When we teach sorting, we take into account in-core vs. out-of-core, the cost of doing a comparison, and the cost of copying the data versus using indices. Different sort algorithms work better with different dataset characteristics. Similarly, we must teach a range of parallel algorithms, so a programmer knows what dataset and system characteristics will affect the performance, and hence the choice of algorithm.

Second, we need to expand algorithm analysis to include different parallelism styles. It’s not enough to focus on just the BSP or SIMD or any other model; we must understand several models and how they map onto the target systems. We’ve become lazy; the sequential von Neumann model has been sufficiently universal that we think we can use a single model to cover all computing. Yes, this sounds like work, and it sounds like we’re going to have to think for a living.

Finally, we need to learn how to analyze and tune actual parallel programs. Sequential performance tuning has largely been reduced to (at most) profiling to find the computationally-intensive region of code, and either choosing a more efficient algorithm, choosing a different set of compiler flags or different compiler, or using a canned library routine. Advanced analysis may consider cache bandwidth and interference or operating system jitter. Parallel performance analysis is subject to many more pitfalls. Threads or processes may be delayed due to synchronization or communication delays; shared cache usage may cause additional interference; NUMA memory accesses may require more careful attention to memory allocation strategies. Program analysis tools can help here.

Stanford’s Computer Science Department Chair, Bill Dally, is quoted as saying “parallel programming is perhaps the largest problem in computer science today and is the major obstacle to the continued scaling of computing performance that has fueled the computing industry, and several related industries, for the last 40 years.” David Patterson, former ACM president and head of the Parallel Computing Laboratory at Cal Berkeley, said that “in order for parallelism to succeed, it has to result in better productivity, efficiency, and accuracy.”

Parallel processing is an obstacle, but then so is sequential processing. Parallel computing can result in better productivity, efficiency, and accuracy in the scientific process overall, but it’s silly to think that it will result in better productivity and efficiency in the programming process itself. The best we can hope for is to make parallel programming not much harder than sequential programming. Dally himself is giving a keynote speech at the International Conference on Parallel Processing in Portland in September titled “Streams: Parallel Programming Made Simple.” There’s that simple word yet again.

That’s not to say there aren’t problems to be solved. There are, and parallel programming is going to continue to be a problem. However, unlike the doomsayers and the simplifiers, I think these are hard problems, but not impossible. We solve hard problems every day. We’ve already developed several successful, widely-used parallel programming paradigms, including MPI and OpenMP. They may not be perfect, universal models, but we should learn what worked. And we have as much to learn from proposed models that did not succeed or survive, such as High Performance Fortran.

And, frankly, I have confidence in the applications programmer’s ability to develop algorithms and approaches to using parallelism. Apparently, many in the computer science community do not.

—–

Michael Wolfe has developed compilers for over 30 years in both academia and industry, and is now a senior compiler engineer at The Portland Group, Inc. (www.pgroup.com), a wholly-owned subsidiary of STMicroelectronics, Inc. The opinions stated here are those of the author, and do not represent opinions of The Portland Group, Inc. or STMicroelectronics, Inc.

Subscribe to HPCwire's Weekly Update!

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

PRACEdays Reflects Europe’s HPC Commitment

May 25, 2017

More than 250 attendees and participants came together for PRACEdays17 in Barcelona last week, part of the European HPC Summit Week 2017, held May 15-19 at t Read more…

By Tiffany Trader

Russian Researchers Claim First Quantum-Safe Blockchain

May 25, 2017

The Russian Quantum Center today announced it has overcome the threat of quantum cryptography by creating the first quantum-safe blockchain, securing cryptocurr Read more…

By Doug Black

Google Debuts TPU v2 and will Add to Google Cloud

May 25, 2017

Not long after stirring attention in the deep learning/AI community by revealing the details of its Tensor Processing Unit (TPU), Google last week announced the Read more…

By John Russell

Nvidia CEO Predicts AI ‘Cambrian Explosion’

May 25, 2017

The processing power and cloud access to developer tools used to train machine-learning models are making artificial intelligence ubiquitous across computing pl Read more…

By George Leopold

HPE Extreme Performance Solutions

Exploring the Three Models of Remote Visualization

The explosion of data and advancement of digital technologies are dramatically changing the way many companies do business. With the help of high performance computing (HPC) solutions and data analytics platforms, manufacturers are developing products faster, healthcare providers are improving patient care, and energy companies are improving planning, exploration, and production. Read more…

PGAS Use will Rise on New H/W Trends, Says Reinders

May 25, 2017

If you have not already tried using PGAS, it is time to consider adding PGAS to the programming techniques you know. Partitioned Global Array Space, commonly kn Read more…

By James Reinders

Exascale Escapes 2018 Budget Axe; Rest of Science Suffers

May 23, 2017

President Trump's proposed $4.1 trillion FY 2018 budget is good for U.S. exascale computing development, but grim for the rest of science and technology spend Read more…

By Tiffany Trader

Hedge Funds (with Supercomputing help) Rank First Among Investors

May 22, 2017

In case you didn’t know, The Quants Run Wall Street Now, or so says a headline in today’s Wall Street Journal. Quant-run hedge funds now control the largest Read more…

By John Russell

IBM, D-Wave Report Quantum Computing Advances

May 18, 2017

IBM said this week it has built and tested a pair of quantum computing processors, including a prototype of a commercial version. That progress follows an an Read more…

By George Leopold

PRACEdays Reflects Europe’s HPC Commitment

May 25, 2017

More than 250 attendees and participants came together for PRACEdays17 in Barcelona last week, part of the European HPC Summit Week 2017, held May 15-19 at t Read more…

By Tiffany Trader

PGAS Use will Rise on New H/W Trends, Says Reinders

May 25, 2017

If you have not already tried using PGAS, it is time to consider adding PGAS to the programming techniques you know. Partitioned Global Array Space, commonly kn Read more…

By James Reinders

Exascale Escapes 2018 Budget Axe; Rest of Science Suffers

May 23, 2017

President Trump's proposed $4.1 trillion FY 2018 budget is good for U.S. exascale computing development, but grim for the rest of science and technology spend Read more…

By Tiffany Trader

Cray Offers Supercomputing as a Service, Targets Biotechs First

May 16, 2017

Leading supercomputer vendor Cray and datacenter/cloud provider the Markley Group today announced plans to jointly deliver supercomputing as a service. The init Read more…

By John Russell

HPE’s Memory-centric The Machine Coming into View, Opens ARMs to 3rd-party Developers

May 16, 2017

Announced three years ago, HPE’s The Machine is said to be the largest R&D program in the venerable company’s history, one that could be progressing tow Read more…

By Doug Black

What’s Up with Hyperion as It Transitions From IDC?

May 15, 2017

If you’re wondering what’s happening with Hyperion Research – formerly the IDC HPC group – apparently you are not alone, says Steve Conway, now senior V Read more…

By John Russell

Nvidia’s Mammoth Volta GPU Aims High for AI, HPC

May 10, 2017

At Nvidia's GPU Technology Conference (GTC17) in San Jose, Calif., this morning, CEO Jensen Huang announced the company's much-anticipated Volta architecture a Read more…

By Tiffany Trader

HPE Launches Servers, Services, and Collaboration at GTC

May 10, 2017

Hewlett Packard Enterprise (HPE) today launched a new liquid cooled GPU-driven Apollo platform based on SGI ICE architecture, a new collaboration with NVIDIA, a Read more…

By John Russell

Quantum Bits: D-Wave and VW; Google Quantum Lab; IBM Expands Access

March 21, 2017

For a technology that’s usually characterized as far off and in a distant galaxy, quantum computing has been steadily picking up steam. Just how close real-wo Read more…

By John Russell

Trump Budget Targets NIH, DOE, and EPA; No Mention of NSF

March 16, 2017

President Trump’s proposed U.S. fiscal 2018 budget issued today sharply cuts science spending while bolstering military spending as he promised during the cam Read more…

By John Russell

Google Pulls Back the Covers on Its First Machine Learning Chip

April 6, 2017

This week Google released a report detailing the design and performance characteristics of the Tensor Processing Unit (TPU), its custom ASIC for the inference Read more…

By Tiffany Trader

HPC Compiler Company PathScale Seeks Life Raft

March 23, 2017

HPCwire has learned that HPC compiler company PathScale has fallen on difficult times and is asking the community for help or actively seeking a buyer for its a Read more…

By Tiffany Trader

CPU-based Visualization Positions for Exascale Supercomputing

March 16, 2017

Since our first formal product releases of OSPRay and OpenSWR libraries in 2016, CPU-based Software Defined Visualization (SDVis) has achieved wide-spread adopt Read more…

By Jim Jeffers, Principal Engineer and Engineering Leader, Intel

Nvidia Responds to Google TPU Benchmarking

April 10, 2017

Last week, Google reported that its custom ASIC Tensor Processing Unit (TPU) was 15-30x faster for inferencing workloads than Nvidia's K80 GPU (see our coverage Read more…

By Tiffany Trader

Nvidia’s Mammoth Volta GPU Aims High for AI, HPC

May 10, 2017

At Nvidia's GPU Technology Conference (GTC17) in San Jose, Calif., this morning, CEO Jensen Huang announced the company's much-anticipated Volta architecture a Read more…

By Tiffany Trader

TSUBAME3.0 Points to Future HPE Pascal-NVLink-OPA Server

February 17, 2017

Since our initial coverage of the TSUBAME3.0 supercomputer yesterday, more details have come to light on this innovative project. Of particular interest is a ne Read more…

By Tiffany Trader

Leading Solution Providers

Facebook Open Sources Caffe2; Nvidia, Intel Rush to Optimize

April 18, 2017

From its F8 developer conference in San Jose, Calif., today, Facebook announced Caffe2, a new open-source, cross-platform framework for deep learning. Caffe2 is Read more…

By Tiffany Trader

Tokyo Tech’s TSUBAME3.0 Will Be First HPE-SGI Super

February 16, 2017

In a press event Friday afternoon local time in Japan, Tokyo Institute of Technology (Tokyo Tech) announced its plans for the TSUBAME3.0 supercomputer, which w Read more…

By Tiffany Trader

Is Liquid Cooling Ready to Go Mainstream?

February 13, 2017

Lost in the frenzy of SC16 was a substantial rise in the number of vendors showing server oriented liquid cooling technologies. Three decades ago liquid cooling Read more…

By Steve Campbell

MIT Mathematician Spins Up 220,000-Core Google Compute Cluster

April 21, 2017

On Thursday, Google announced that MIT math professor and computational number theorist Andrew V. Sutherland had set a record for the largest Google Compute Eng Read more…

By Tiffany Trader

US Supercomputing Leaders Tackle the China Question

March 15, 2017

As China continues to prove its supercomputing mettle via the Top500 list and the forward march of its ambitious plans to stand up an exascale machine by 2020, Read more…

By Tiffany Trader

HPC Technique Propels Deep Learning at Scale

February 21, 2017

Researchers from Baidu's Silicon Valley AI Lab (SVAIL) have adapted a well-known HPC communication technique to boost the speed and scale of their neural networ Read more…

By Tiffany Trader

DOE Supercomputer Achieves Record 45-Qubit Quantum Simulation

April 13, 2017

In order to simulate larger and larger quantum systems and usher in an age of "quantum supremacy," researchers are stretching the limits of today's most advance Read more…

By Tiffany Trader

Knights Landing Processor with Omni-Path Makes Cloud Debut

April 18, 2017

HPC cloud specialist Rescale is partnering with Intel and HPC resource provider R Systems to offer first-ever cloud access to Xeon Phi "Knights Landing" process Read more…

By Tiffany Trader

  • arrow
  • Click Here for More Headlines
  • arrow
Share This