Visit additional Tabor Communication Publications
February 23, 2012
The most widely used computer programming languages today were not designed as parallel programming languages. But retrofitting existing programming languages for parallel programming is underway. We can compare and contrast retrofits by looking at four key features, five key qualities, and the various implementation approaches.
In this article, I focus on the features and qualities, leaving the furious debates over best approaches (language vs. library vs. directives, and abstract and portable vs. low-level with lots of controls) for another day.
Four features we need
Features that any parallel programming solution, including retrofits, should include a defined memory model, synchronization, tasks, and data parallelism.
Defining how changes in shared data are observable by different tasks had been an under-appreciated problem. Hans-J. Boehm wrote a report in 2004, titled Threads Cannot Be Implemented As a Library, which explains these issues. Having a well-defined ordering among accesses to distinct variables, and enabling the independence of updates to distinct variables, is so important that they have been addressed in Java, C11 and C++11. Without these retrofits, every parallel program sits on a crumbling foundation.
The need for portable and efficient synchronization is substantial. Boost libraries, Intel’s Threading Building Blocks (TBB) and OpenMP offer solutions that are widely utilized. C++11 and C11 now offer support. Beyond these, the concept of transactions is a topic worth exploring in a future article. Synchronization retrofitting is helping portability. Substantial opportunities remain for helping efficiency.
Tasks, not threads
Programming should be an exercise in writing tasks that can run concurrently, without the programmer specifying the precise mapping of tasks onto hardware threads. An introduction to this challenge is The Problem with Threads by Edward A. Lee.
Mapping should be the job of tools, including run-time schedulers, not explicit programming. This philosophy is being well supported by retrofits like OpenMP, TBB, Cilk Plus, Microsoft’s Parallel Patterns Library (PPL) and Apple’s Grand Central Dispatch (GCD). The need to assert some control over task to thread mapping to maximize performance is still present when using such systems today, but not always supported.
Nevertheless, programming directly to native threads (e.g., pthreads) in applications is something that should be completely avoided. Retrofits are sufficient today to make tasks the method of choice.
Data parallel support
It should be reasonably straightforward to write a portable program that takes advantage of data parallel hardware. Ideally, data parallel support should be able to utilize vector and task parallel capabilities without a programmer having to explicitly code the division between the two.
Unfortunately, no such solution is in wide spread use today even for vectorization alone. Effective auto-parallelization is very dependent on highly optimizing compilers. Compiler intrinsics lock code into a particular vector width (MMX=64, SSE=128, AVX=256, etc.). Elemental functions in CUDA, OpenCL, and Cilk Plus offer a glimpse into possible retrofits. Intel proposes we adopt the vectorization benefits of Fortran 90 array notations into C and C++ as part of the Cilk Plus project.
Vector hardware is increasingly important in processors, GPUs and co-processors. OpenCL and OpenMP wrestle today with how to embrace data parallel hardware and how tightly tied programming will be to it. Microsoft C++ AMP has similar challenges when it comes to market with the next Microsoft Visual Studio. Standard, abstract, portable and effective solutions wanted!
Five qualities we should desire
Five key qualities that are desirable, for parallel programming, include composability, sequential reasoning, communication minimization, performance portability and safety.
All of these qualities are unobtainable, in an absolute sense, whether as retrofits in an old language or with a clean slate and a new language. That is why we cannot call them features. The more of these qualities we obtain the better off we are. That makes them very important to keep in mind.
Composability is a well-known concept in programming, offering rules for combining different things together (functions, objects, modules, etc.) so that it is easy to compose (think: combine in unanticipated ways). It is important to think of composability in terms of both correctness and performance.
OpenCL, largely because it is less abstract, has low composability on both accounts. OpenMP and OpenCL have very serious performance composability unless they are used very carefully. New and abstract retrofits (TBB, Cilk, PPL, GCD) are much more tolerant and able to deliver high composability.
Self-composability is an essential first step, but the ability to compose multiple retrofits together is essential in the long run as well. A welcome solution for tool vendors, Microsoft’s Concurrency Runtime has allowed retrofits from multiple vendors to coexist with increased composability. Parallel programming without the ability to mix and match freely, is undesirable and counterproductive.
Composability deserves more attention than it gets.
Sequential reasoning, the norm for reading a serial implementation, can apply with an appropriately expressed parallel program. OpenMP uses hints to create the use of parallelism instead of code changes. This allows the intent of a program to remain evident in the code. TBB and PPL emphasize relaxed sequential semantics to provide parallelism as an accelerator without making it mandatory for correctness. Writing a program in a sequentially consistent fashion is permitted and encouraged.
An explicit goal of Cilk Plus is to offer sequential semantics to set it apart from other retrofits. The serial elision (or C elision) of a Cilk program is touted in papers from MIT. Programming that preserves sequential semantics has received praise as easier to learn and use. The elemental functions in OpenCL, CUDA and Cilk Plus have similar objectives.
It is fair to say that programming in a manner that requires understanding parallel semantics, in order to understand intent, is both unpopular and out of vogue today. Such mandatory parallelism is harder to understand and to debug.
Sequential reasoning can be extended to debuggers too. A hot area to watch here is debuggers working to present a debugging experience more akin to sequential experiences, with features like Rogue Wave’s replay capabilities in the Totalview debugger.
Instead of sequential reasoning being a retrofit, it is more accurate to think of sequential reasoning as often being purposefully sought and preserved in a parallel world.
Performance tuning on parallel systems often focuses on ensuring data is local when you use it and minimizing the need to move it around. Data motion means communication of some sort, and communication is generally expensive. Decisions in the design and implementation of retrofits, as well as the application programming itself, often impact performance dramatically. The task stealing algorithms of TBB, Cilk, PPL and GCD all have cache reuse strongly in mind in their designs. Retrofits to help, with communication minimization, are a tricky business and could use more attention.
The goal here is that a tuned program on one piece of hardware performs reasonably well on another piece of hardware. It is desirable to be able to describe data and tasks in such a way that performance scales as parallelism increases (number of cores, or size of vectors, or cache size, etc.). Nothing is ever fully performance portable, but more abstract retrofits tend to hold up better.
Unfortunately, implementations of abstractions can struggle to offer peak performance. It took years for compilers to offer performance for MMX or SSE that was competitive with assembly language programming. Use of cache-agnostic algorithms generally increase performance portability. Today, competing on performance with carefully-crafted CUDA and OpenCL code can be challenging because the coding is low level enough to encourage, or even require, the program structure to match the hardware. The lack of performance portability of such code is frequently shown, but effective alternatives remain works-in-progress. Language design, algorithm choices and programming style can affect performance portability a great deal.
The freedom from deadlocks and race conditions, may be the most difficult to provide via a retrofit. No method to add complete safety to C or C++ has gained wide popularity. Safety has not been incorporated into non-managed languages easily, despite some valiant efforts to do so.
To make a language safe, pointers have to be removed or severely restricted. Meanwhile, tools are maturing to help us cope with safety despite lack of direct language support, and safer coding style and safer retrofits appear to help as well. Perhaps safety comes via a combination of “good enough” and “we can cope using tools.”
A journey ahead, together
There are at least four key programming problems that any parallel programming solution should address, and five key qualities that can make a programming model, retrofit or otherwise, more desirable. Evolution in hardware will help as well.
About the author
James Reinders has helped develop supercomputers, microprocessors and software tools for 25 years. He is a senior engineer for Intel in Hillsboro Oregon.
Jun 19, 2013 |
Supercomputer architectures have evolved considerably over the last 20 years, particularly in the number of processors that are linked together. One aspect of HPC architecture that hasn't changed is the MPI programming model.
Jun 18, 2013 |
The world's largest supercomputers, like Tianhe-2, are great at traditional, compute-intensive HPC workloads, such as simulating atomic decay or modeling tornados. But data-intensive applications--such as mining big data sets for connections--is a different sort of workload, and runs best on a different sort of computer.
Jun 18, 2013 |
Researchers are finding innovative uses for Gordon, the 285 teraflop supercomputer housed at the San Diego Supercomputer Center (SDSC) that has a unique Flash-based storage system. Since going online, researchers have put the incredibly fast I/O to use on a wide variety of workloads, ranging from chemistry to political science.
Jun 17, 2013 |
The advent of low-power mobile processors and cloud delivery models is changing the economics of computing. But just as an economy car is good at different things than a full size truck, an HPC workload still has certain computing demands that neither the fastest smartphone nor the most elastic cloud cluster can fulfill.
Jun 14, 2013 |
For all the progress we've made in IT over the last 50 years, there's one area of life that has steadfastly eluded the grasp of computers: understanding human language. Now, researchers at the Texas Advanced Computing Center (TACC) are utilizing a Hadoop cluster on its Longhorn supercomputer to move the state of the art of language processing a little bit further.
05/10/2013 | Cleversafe, Cray, DDN, NetApp, & Panasas | From Wall Street to Hollywood, drug discovery to homeland security, companies and organizations of all sizes and stripes are coming face to face with the challenges – and opportunities – afforded by Big Data. Before anyone can utilize these extraordinary data repositories, however, they must first harness and manage their data stores, and do so utilizing technologies that underscore affordability, security, and scalability.
04/15/2013 | Bull | “50% of HPC users say their largest jobs scale to 120 cores or less.” How about yours? Are your codes ready to take advantage of today’s and tomorrow’s ultra-parallel HPC systems? Download this White Paper by Analysts Intersect360 Research to see what Bull and Intel’s Center for Excellence in Parallel Programming can do for your codes.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?
Join our webinar to learn how IT managers can migrate to a more resilient, flexible and scalable solution that grows with the data center. Mellanox VMS is future-proof, efficient and brings significant CAPEX and OPEX savings. The VMS is available today.