Many of the latest supercomputers are based on accelerators, including the two fastest systems according to the 11/2013 TOP500 list. Accelerators are also becoming widespread in PCs and are even starting to appear in handheld devices, which will further boost the interest in accelerator programming.
This broad adoption is the result of high performance, good energy efficiency, and low price. For example, comparing a Xeon E5-2687W CPU to a GTX 680 GPU, both of which were released in March 2012, we find that the GPU started out four times cheaper, has eight times more single-precision performance and four times more main-memory bandwidth, and provides over thirty times as much performance per dollar and six times as much performance per watt. Based on these numbers, accelerators should be used everywhere and all the time. So why aren’t they?
There are two main difficulties with accelerators. First, they can only execute certain types of programs efficiently, in particular programs with sufficient parallelism, data reuse, and regularity in their control flow and memory access patterns. Second, it is harder to write effective software for accelerators than for CPUs because of architectural disparities such as very wide parallelism, exposed memory hierarchies, lockstep execution, and memory-access coalescing. Several new programming languages and extensions thereof have been proposed to hide these aspects to various degrees and thus make it easier to program accelerators.
The initial attempts to use GPUs, which are currently the most prominent type of accelerator, for speeding up non-graphics programs were cumbersome and required expressing the computation in form of shader code that only supported limited control flow and no integer operations. Gradually, these constraints were lifted, making GPUs more general-purpose computing devices and enabling non-graphics-experts to program them. The biggest step in this direction came with the release of the CUDA programming language. It extends C/C++ with additional qualifiers and keywords as well as library functions and a mechanism to launch code sections, called kernels, on a GPU.
The rapid adoption of CUDA, combined with that fact that it is proprietary and the complexity of writing good CUDA code, triggered the creation of several other programming approaches for accelerators, including OpenCL, C++ AMP, and OpenACC. OpenCL is the non-proprietary counterpart of CUDA and is backed by many large companies. It is not restricted to NVIDIA GPUs but also supports AMD GPUs, multicore CPUs, MICs (Intel’s Xeon Phi), DSPs, and FPGAs, making it very portable. However, just like CUDA, it is very low level and requires the software developer to explicitly orchestrate data movement, select where variables live in the memory hierarchy, and manually express parallelism in the code. C++ Accelerated Massive Parallelism (C++ AMP) operates at a medium level. It allows expressing data parallelism directly in C++ and hides all low-level code from the programmer. Parallel “for each” statements encapsulate parallel code. C++ AMP is tied to Windows, does not (yet) support CPUs, and suffers from startup overhead, making it impractical for accelerating short-running code sections.
OpenACC is a very high-level approach that allows programmers to annotate their code using pragmas to inform the compiler which code sections to accelerate, e.g., by offloading them to a GPU. The idea is similar to how OpenMP can be used to parallelize CPU programs. In fact, there are efforts underway to merge the two approaches. OpenACC is still maturing and currently only supported by a few compilers.
To predict how accelerator programming might develop from this point forward, it may be helpful to study how other acceleration hardware has evolved in the past. For example, some early high-end PCs contained an extra chip, called a co-processor, to accelerate floating-point (FP) calculations. Later, this co-processor was combined with the CPU on the same die and is now fully integrated with the integer processing core. Only separate FP registers and ALUs remain. The much more recently added SIMD support (including MMX, SSE, AltiVec, and AVX) did not start out on a separate chip but is now also fully integrated in the core. Just like the floating-point instructions, SIMD instructions operate on separate registers and ALUs.
Interestingly, the programmer’s view of these two types of instructions is surprisingly different. The floating-point operations and data types have been standardized long ago (IEEE 754) and are now ubiquitous. They are directly available in high-level languages through normal arithmetic operations and built-in 32-bit single-precision and 64-bit double-precision types. In contrast, no standard exists for SIMD instructions, and their existence is largely hidden from programmers. It is left to the compiler to ‘vectorize’ the code and employ these instructions. Developers wishing to use SIMD instructions explicitly have to resort to compiler-specific macros that are not portable.
Since GPUs and MICs obtain their high performance through SIMD-like execution, we believe accelerators are more likely to track the evolution of SIMD- than FP-instruction support. Another similarity to SIMD, and a key factor that made CUDA so successful, is that CUDA hides the SIMD aspect of the GPU hardware and allows the programmer to think in terms of individual threads operating on scalar data elements rather than warps operating on vectors. Hence, accelerators will undoubtedly also be moved onto the CPU chip, but we surmise that their code will not be seamlessly interwoven with CPU code nor will the accelerators’ hardware-supported data types be made explicitly available to the programmers.
Some accelerators have already been combined with conventional processing cores on the same chip, including on AMD’s APUs (as used in the Xbox One), Intel’s processors with HD Graphics, and NVIDIA’s Tegra SoC. However, accelerators will probably remain separate cores because it is difficult to fuse accelerator and conventional cores to the degree that was possible with FP and SIMD instructions, i.e., to ‘reduce’ the accelerator to just a set of separate registers and ALUs in the general-purpose core. After all, accelerators are so fast, parallel, and energy efficient because they are based on dissimilar architectural tradeoffs, such as incoherent caches, very different pipeline designs, GDDR5 memory, and an order of magnitude more registers and multithreading. Hence, the complexity of having to run separate accelerator code will remain. As even cores on the same die tend to share no more than the bottom of the memory hierarchy, data transfers between CPU and accelerator cores will possibly become faster but also remain a bottleneck.
The explicit orchestration of data exchanges between devices is a significant source of errors and a substantial burden on programmers. For short kernels, it is often the case that more code needs to be written to transfer data back and forth than to express the actual computation. Eliminating this burden is one of the primary benefits of higher-level programming approaches such as C++ AMP and OpenACC. Even low-level approaches have been addressing this problem. For example, streamlined and unified memory addressing is one of the major improvements in the latest CUDA and OpenCL releases and NVIDIA GPU hardware. Yet, to achieve good performance, some help by the programmer is generally needed, even in very high-level approaches like OpenACC. In particular, locality-aware memory allocation and data migration often have to be handled manually.
Unfortunately, any ease provided by such improvements may only turn out to be a partial solution. Based on the assumption that future microprocessors will be similar to today’s (small) supercomputers, it is likely that they will contain many more cores than can be served by a shared (NUMA) memory system. Instead, we believe there will be clusters of cores on each die where each cluster has its own memory, possibly stacked on top of the cores in a 3D design. The clusters communicate with each other via an on-chip network using a protocol akin to MPI. We do not believe this is farfetched as Intel just announced that it will include networking capabilities in their future Xeon chips, which is a step in this direction. Hence, it is likely that future chips will become more and more heterogeneous, comprising latency- and throughput-optimized cores, NICs, encryption and compression cores, FPGAs, etc.
That raises the all-important question of how to program such devices. We think the answer is surprisingly similar to how today’s multiple CPU cores, SIMD instruction set extensions, and compute accelerators are being used. Basically, there are three levels at which this is done, which we refer to as libraries, automated tools, and do-it-yourself. The library approach is the simplest and works by calling functions in a library that has been accelerated by someone else. Many of the latest math libraries belong to this category. As long as most of the computation takes place inside the library code, this approach is very successful. It allows a few expert library writers to enable the acceleration of a large number of applications.
The automated tools approach is the approach taken by C++ AMP and OpenACC, where the compiler has to do the heavy lifting. The success of this approach depends on the quality and sophistication of the available software tools and, as mentioned, often needs help from the programmer. Nevertheless, most developers can reasonably quickly achieve positive results with this approach, which is not limited to predetermined functions in a library. This is perhaps reminiscent of how a few expert teams code up the inner workings of SQL, which then allows a large number of ‘regular’ programmers to benefit from the optimizations and know-how that the experts encoded.
Finally, the do-it-yourself approach is represented by CUDA and OpenCL, which give the programmer full control over and access to almost every aspect of the accelerator. If implemented well, the resulting code can outperform the other two approaches. However, this comes at the cost of a steep learning curve, lots of extra code that needs to be written, and a slew of additional possibilities for bugs. Ever improving debugging and programming environments will help alleviate these problems, but only to a degree. Hence, this approach is primarily useful for expert programmers, such as those writing the aforementioned libraries and tools.
Since it is trivial to use, programmers will employ the library approach whenever possible. However, this hinges on the availability of appropriate library functions, which are easy to provide in well-defined and mature domains such as standard matrix operations (BLAS) but hard to do for emerging areas or unstructured computations. In the absence of adequate libraries, a programmer’s second choice will be the tools approach, assuming that the tools will be mature. Any computations that are not available in a library, do not demand the highest performance, and are supported by the compiler will likely be coded up using the tools approach. For the remaining cases, the do-it-yourself approach will have to be used. Since OpenCL incorporates the successful ideas that CUDA introduced, is non-proprietary, and supports a large range of hardware, we believe OpenCL or a derivative of it will start dominating this domain akin to how MPI has become the de facto standard for distributed-memory programming.
Taking the union of the hardware features and evolution outlined above, future processor chips might contain multiple clusters with their own memory, each cluster consists of a set of cores where not all cores necessarily have the same capabilities, each multithreaded compute core comprises a large number of processing elements (i.e., functional units or ALUs), and each processing element may be able to perform SIMD operations. Even though actual chips might not include all of the above, they all share a key similarity, namely a hierarchy of distinct parallelization levels. To effectively and portably program such a system, we propose what we call the “copious-parallelism” technique. It is a generalization of how MPI programs are typically explicitly written to adapt to the number of available compute nodes or how OpenMP code implicitly adapts to the number of available cores (or threads).
The main idea behind copious parallelism, and the reason for its name, is to provide ample and parameterizable parallelism for each level. The parameterization makes it possible to decrease the parallelization at any level to match the hardware parallelism at that level. For example, on a shared-memory system, the highest level of parallelization is not necessary and should be set to just one “cluster”. Similarly, in a core where the functional units are unable to perform SIMD instructions, the parameter determining the SIMD width should be set to one. This technique is able to exploit the common features of current multicore CPUs, GPUs, MICs, and other devices as well as likely future architectures. While it is definitely harder to write software in this manner, copious parallelism makes it possible to extract high performance from a broad range of computing devices with just a single code base.
We have tested this approach on a direct n-body simulation [ref]. We wrote a single copious-parallelism implementation in OpenCL and assessed it on four very different architectures: an NVIDIA GeForce Titan GPU, an AMD Radeon 7970 GPU, an Intel Xeon E5-2690 multicore CPU, and an Intel Xeon Phi 5110P MIC. Given our FLOP mix of 54% non-FMA operations, the copious-parallelism code achieves 75% of the theoretical peak performance on the Titan, 95% on the Radeon, 80.5% on the CPU, and 80% on the MIC. While this is only one example, the results are extremely encouraging. In fact, we believe the copious-parallelism technique may well be and remain for quite some time the only portable high-performance approach for programming current and future accelerated systems.
About the Authors
Kamil Rocki is a postdoctoral researcher at IBM Research. Prior to joining Almaden Research Center in California he has spent 5 years in Japan at the University of Tokyo where he graduated with a PhD degree in Information Science. Before that he received his M.Sc. and B.Sc. degrees in Computer Science from Warsaw University of Technology. His current research focuses on high performance parallel algorithms and hardware design. Other interests include supercomputing, AI, computer vision and robotics. He has been working in the field of GPGPU programming for the past 8 years.
Martin Burtscher is Associate Professor in the Department of Computer Science at Texas State University. He received the BS/MS degree in computer science from the Swiss Federal Institute of Technology (ETH) Zurich and the Ph.D. degree in computer science from the University of Colorado at Boulder. Martin’s research interests include efficient parallelization of programs for GPUs and multicore CPUs, automatic performance assessment and optimization, and high-speed data compression. He is a senior member of the IEEE, its Computer Society, and the ACM. Martin has co-authored over 75 peer-reviewed scientific publications, including a book chapter in NVIDIA’s GPU Computing Gems, is the recipient of an NVIDIA Academic Partnership award, and is the PI of a CUDA Teaching Center.