Exploring the interoperability of remote GPGPU virtualization using rCUDA and directive-based programming models

Directive-based programming models, such as OpenMP, OpenACC, and OmpSs, enable users to accelerate applications by using coprocessors with little effort. These devices offer significant computing power, but their use can introduce two problems: an increase in the total cost of ownership and their underutilization because not all codes match their architecture. Remote accelerator virtualization frameworks address those problems. In particular, rCUDA provides transparent access to any graphic processor unit installed in a cluster, reducing the number of accelerators and increasing their utilization ratio. Joining these two technologies, directive-based programming models and rCUDA, is thus highly appealing. In this work, we study the integration of OmpSs and OpenACC with rCUDA, describing and analyzing several applications over three different hardware configurations that include two InfiniBand interconnections and three NVIDIA accelerators. Our evaluation reveals favorable performance results, showing low overhead and similar scaling factors when using remote accelerators instead of local devices.     

Performance of CPU/GPU compiler directives on ISO/TTI kernels

GPUs are slowly becoming ubiquitous devices in High Performance Computing, as their capabilities to enhance the performance per watt of compute intensive algorithms as compared to multicore CPUs have been identified. The primary shortcoming of a GPU is usability, since vendor specific APIs are quite different from existing programming languages, and it requires a substantial knowledge of the device and programming interface to optimize applications. Hence, lately a growing number of higher level programming models are targeting GPUs to alleviate this problem. The ultimate goal for a high-level model is to expose an easy-to-use interface for the user to offload compute intensive portions of code (kernels) to the GPU, and tune the code according to the target accelerator to maximize overall performance with a reduced development effort. In this paper, we share our experiences of three of the notable high-level directive based GPU programming models—PGI, CAPS and OpenACC (from CAPS and PGI) on an Nvidia M2090 GPU. We analyze their performance and programmability against Isotropic (ISO)/Tilted Transversely Isotropic (TTI) finite difference kernels, which are primary components in the Reverse Time Migration (RTM) application used by oil and gas exploration for seismic imaging of the sub-surface. When ported to a single GPU using the mentioned directives, we observe an average 1.5–1.8x improvement in performance for both ISO and TTI kernels, when compared with optimized multi-threaded CPU implementations using OpenMP.     

Accelerating the SRP-PHAT algorithm on multi- and many-core platforms using OpenCL

The Steered Response Power with Phase Transform (SRP-PHAT) algorithm is a well-known method for sound source localization due to its robust performance in noisy and reverberant environments. This algorithm is used in a large number of acoustic applications such as automatic camera steering systems, human–machine interaction, video gaming and audio surveillance. SPR-PHAT implementations require to handle a high number of signals coming from a microphone array and a huge search grid that influences the localization accuracy of the system. In this context, high performance in the localization process can only be achieved by using massively parallel computational resources. Different types of multi-core machines based either on multiple CPUs or on GPUs are commonly employed in diverse fields of science for accelerating a number of applications, mainly using OpenMP and CUDA as programming frameworks, respectively. This implies the development of multiple source codes which limits the portability and application possibilities. On the contrary, OpenCL has emerged as an open standard for parallel programming that is nowadays supported by a wide range of architectures. In this work, we evaluate an OpenCL-based implementations of the SRP-PHAT algorithm in two state-of-the-art CPU and GPU platforms. Results demonstrate that OpenCL achieves close-to-CUDA performance in GPU (considered as upper bound) and outperforms in most of the CPU configurations based on OpenMP.

     

Data Parallel Algorithmic Skeletons with Accelerator Support

Hardware accelerators such as GPUs or Intel Xeon Phi comprise hundreds or thousands of cores on a single chip and promise to deliver high performance. They are widely used to boost the performance of highly parallel applications. However, because of their diverging architectures programmers are facing diverging programming paradigms. Programmers also have to deal with low-level concepts of parallel programming that make it a cumbersome task. In order to assist programmers in developing parallel applications Algorithmic Skeletons have been proposed. They encapsulate well-defined, frequently recurring parallel programming patterns, thereby shielding programmers from low-level aspects of parallel programming. The main contribution of this paper is a comparison of two skeleton library implementations, one in C++ and one in Java, in terms of library design and programmability. Besides, on the basis of four benchmark applications we evaluate the performance of the presented implementations on two test systems, a GPU cluster and a Xeon Phi system. The two implementations achieve comparable performance with a slight advantage for the C++ implementation. Xeon Phi performance ranges between CPU and GPU performance.     

DaSH: A benchmark suite for hybrid dataflow and shared memory programming models

The current trend in development of parallel programming models is to combine different well established models into a single programming model in order to support efficient implementation of a wide range of real world applications. The dataflow model has particularly managed to recapture the interest of the research community due to its ability to express parallelism efficiently. Thus, a number of recently proposed hybrid parallel programming models combine dataflow and traditional shared memory models. Their findings have influenced the introduction of task dependency in the OpenMP 4.0 standard.This article presents DaSH – the first comprehensive benchmark suite for hybrid dataflow and shared memory programming models. DaSH features 11 benchmarks, each representing one of the Berkeley dwarfs that capture patterns of communication and computation common to a wide range of emerging applications. DaSH also includes sequential and shared-memory implementations based on OpenMP and Intel TBB to facilitate easy comparison between hybrid dataflow implementations and traditional shared memory implementations based on work-sharing and/or tasks. Finally, we use DaSH to evaluate three different hybrid dataflow models, identify their advantages and shortcomings, and motivate further research on their characteristics.     

MPtostream:an OpenMP compiler for CPU-GPU heterogeneous parallel systems

In light of GPUs’ powerful floating-point operation capacity,heterogeneous parallel systems incorporating general purpose CPUs and GPUs have become a highlight in the research field of high performance computing(HPC).However,due to the complexity of programming on GPUs,porting a large number of existing scientific computing applications to the heterogeneous parallel systems remains a big challenge.The OpenMP programming interface is widely adopted on multi-core CPUs in the field of scientific computing.To effectively inherit existing OpenMP applications and reduce the transplant cost,we extend OpenMP with a group of compiler directives,which explicitly divide tasks among the CPU and the GPU,and map time-consuming computing fragments to run on the GPU,thus dramatically simplifying the transplantation.We have designed and implemented MPtoStream,a compiler of the extended OpenMP for AMD’s stream processing GPUs.Our experimental results show that programming with the extended directives deviates from programming with OpenMP by less than 11% modification and achieves significant speedup ranging from 3.1 to 17.3 on a heterogeneous system,incorporating an Intel Xeon E5405 CPU and an AMD FireStream 9250 GPU,over the execution on the Xeon CPU alone.     

多核CPU和GPU加速分子动力学模拟

在多核中央处理器(CPU)—图形处理器(GPU)异构并行体系结构上,采用OpenMP和计算统一设备架构(CUDA)编程实现了基于AMBER力场的蛋白质分子动力学模拟程序。通过合理地将程序划分为CPU单线程、CPU多线程和GPU多线程执行部分,高效地利用了计算机的处理能力。性能测试结果表明,相对于优化后的CPU串行计算,多核CPU-GPU异构并行计算模型有强大的性能优势,特别是将占整个程序执行时间90%的作用力的计算移植到GPU上执行,获得了最高可达12倍的计算加速比。     

A performance study of general-purpose applications on graphics processors using CUDA

Graphics processors (GPUs) provide a vast number of simple, data-parallel, deeply multithreaded cores and high memory bandwidths. GPU architectures are becoming increasingly programmable, offering the potential for dramatic speedups for a variety of general-purpose applications compared to contemporary general-purpose processors (CPUs). This paper uses NVIDIA’s C-like CUDA language and an engineering sample of their recently introduced GTX 260 GPU to explore the effectiveness of GPUs for a variety of application types, and describes some specific coding idioms that improve their performance on the GPU. GPU performance is compared to both single-core and multicore CPU performance, with multicore CPU implementations written using OpenMP. The paper also discusses advantages and inefficiencies of the CUDA programming model and some desirable features that might allow for greater ease of use and also more readily support a larger body of applications.     

Parallelization of 2D MPDATA EULAG algorithm on hybrid architectures with GPU accelerators

EULAG (Eulerian/semi-Lagrangian fluid solver) is an established computational model developed for simulating thermo-fluid flows across a wide range of scales and physical scenarios. The dynamic core of EULAG includes the multidimensional positive definite advection transport algorithm (MPDATA) and elliptic solver. In this work we investigate aspects of an optimal parallel version of the 2D MPDATA algorithm on modern hybrid architectures with GPU accelerators, where computations are distributed across both GPU and CPU components.Using the hybrid OpenMP–OpenCL model of parallel programming opens the way to harness the power of CPU–GPU platforms in a portable way. In order to better utilize features of such computing platforms, comprehensive adaptations of MPDATA computations to hybrid architectures are proposed. These adaptations are based on efficient strategies for memory and computing resource management, which allow us to ease memory and communication bounds, and better exploit the theoretical floating point efficiency of CPU–GPU platforms. The main contributions of the paper are:

• •method for the decomposition of the 2D MPDATA algorithm as a tool to adapt MPDATA computations to hybrid architectures with GPU accelerators by minimizing communication and synchronization between CPU and GPU components at the cost of additional computations;
• •method for the adaptation of 2D MPDATA computations to multicore CPU platforms, based on space and temporal blocking techniques;
• •method for the adaptation of the 2D MPDATA algorithm to GPU architectures, based on a hierarchical decomposition strategy across data and computation domains, with support provided by the developed GPU task scheduler allowing for the flexible management of available resources;
• •approach to the parametric optimization of 2D MPDATA computations on GPUs using the autotuning technique, which allows us to provide a portable implementation methodology across a variety of GPUs.

Hybrid platforms tested in this study contain different numbers of CPUs and GPUs – from solutions consisting of a single CPU and a single GPU to the most elaborate configuration containing two CPUs and two GPUs. Processors of different vendors are employed in these systems – both Intel and AMD CPUs, as well as GPUs from NVIDIA and AMD. For all the grid sizes and for all the tested platforms, the hybrid version with computations spread across CPU and GPU components allows us to achieve the highest performance. In particular, for the largest MPDATA grids used in our experiments, the speedups of the hybrid versions over GPU and CPU versions vary from 1.30 to 1.69, and from 1.95 to 2.25, respectively.     

Hybrid CUDA, OpenMP, and MPI parallel programming on multicore GPU clusters

Nowadays, NVIDIA's CUDA is a general purpose scalable parallel programming model for writing highly parallel applications. It provides several key abstractions – a hierarchy of thread blocks, shared memory, and barrier synchronization. This model has proven quite successful at programming multithreaded many core GPUs and scales transparently to hundreds of cores: scientists throughout industry and academia are already using CUDA to achieve dramatic speedups on production and research codes. In this paper, we propose a parallel programming approach using hybrid CUDA OpenMP, and MPI programming, which partition loop iterations according to the number of C1060 GPU nodes in a GPU cluster which consists of one C1060 and one S1070. Loop iterations assigned to one MPI process are processed in parallel by CUDA run by the processor cores in the same computational node.     

PFACC: An OpenACC-like programming model for irregular nested parallelism

OpenACC is a directive-based programming model which allows programmers to write graphic processing unit (GPU) programs by simply annotating parallel loops. However, OpenACC has poor support for irregular nested parallel loops, which are natural choices to express nested parallelism. We propose PFACC, a programming model similar to OpenACC. PFACC directives can be used to annotate parallel loops and to guide data movement between different levels of memory hierarchy. Parallel loops can be arbitrarily nested or be placed inside functions that would be (possibly recursively) called in other parallel loops. The PFACC translator translates C programs with PFACC directives into CUDA programs by inserting runtime iteration-sharing and memory allocation routines. The PFACC runtime iteration-sharing routine is a two-level mechanism. Thread blocks dynamically organize loop iterations into batches and execute the batches in a depth-first order. Different thread blocks share iterations among one another with an iteration-stealing mechanism. PFACC generates CUDA programs with reasonable memory usage because of the depth-first execution order. The two-level iteration-sharing mechanism is implemented purely in software and fits well with the CUDA thread hierarchy. Experiments show that PFACC outperforms CUDA dynamic parallelism in terms of performance and code size on most benchmarks.     

A comparative performance study of common and popular task‐centric programming frameworks

Programmers today face a bewildering array of parallel programming models and tools, making it difficult to choose an appropriate one for each application. An increasingly popular programming model supporting structured parallel programming patterns in a portable and composable manner is the task‐centric programming model. In this study, we compare several popular task‐centric programming frameworks, including Cilk Plus, Threading Building Blocks, and various implementations of OpenMP 3.0. We have analyzed their performance on the Barcelona OpenMP Tasking Suite benchmark suite both on a 48‐core AMD Opteron 6172 server and a 64‐core TILEPro64 embedded many‐core processor. Our results show that the OpenMP offers the highest flexibility for programmers, and this flexibility comes to a cost. Frameworks supporting only a specific and more restrictive model, such as Cilk Plus and Threading Building Blocks, are generally more efficient both in terms of performance and energy consumption. However, Intel's implementation of OpenMP tasks performs the best and closest to the specialized run‐time systems. Copyright © 2013 John Wiley & Sons, Ltd.     

应用GPU集群加速计算蛋白质分子场

针对生物化学计算中采用量子化学理论计算蛋白质分子场所带来的巨大计算量的问题,搭建起一个GPU集群系统,用来加速计算基于量子化学的蛋白质分子场.该系统采用消息传递并行编程环境(MPI)连接集群各结点,以开放多线程OpenMP编程标准作为多核CPU编程环境,以CUDA语言作为GPU编程环境,提出并实现了集群系统结点中GPU和多核CPU协同计算的并行加速架构优化设计.在保持较高计算精度的前提下,结合MPI,OpenMP和CUDA混合编程模式,大大提高了系统的计算性能,并对不同体系和规模的蛋白质分子场模拟进行了计算分析.与相应的CPU集群、GPU单机和CPU单机计算方法对比,该GPU集群大幅度地提高了高分辨率复杂蛋白质分子场模拟的计算效率,比CPU集群的平均计算加速比提高了7.5倍.     

Performance Evaluation of Mixed-Mode OpenMP/MPI Implementations

With the current prevalence of multi-core processors in HPC architectures mixed-mode programming, using both MPI and OpenMP in the same application, is seen as an important technique for achieving high levels of scalability. As there are few standard benchmarks written in this paradigm, it is difficult to assess the likely performance of such programs. To help address this, we examine the performance of mixed-mode OpenMP/MPI on a number of popular HPC architectures, using a synthetic benchmark suite and two large-scale applications. We find performance characteristics which differ significantly between implementations, and which highlight possible areas for improvement, especially when multiple OpenMP threads communicate simultaneously via MPI.     

Performance comparison of MPI and OpenMP on shared memory multiprocessors

When using a shared memory multiprocessor, the programmer faces the issue of selecting the portable programming model which will provide the best performance. Even if they restricts their choice to the standard programming environments (MPI and OpenMP), they have to select a programming approach among MPI and the variety of OpenMP programming styles. To help the programmer in their decision, we compare MPI with three OpenMP programming styles (loop level, loop level with large parallel sections, SPMD) using a subset of the NAS benchmark (CG, MG, FT, LU), two dataset sizes (A and B), and two shared memory multiprocessors (IBM SP3 NightHawk II, SGI Origin 3800). We have developed the first SPMD OpenMP version of the NAS benchmark and gathered other OpenMP versions from independent sources (PBN, SDSC and RWCP). Experimental results demonstrate that OpenMP provides competitive performance compared with MPI for a large set of experimental conditions. Not surprisingly, the two best OpenMP versions are those requiring the strongest programming effort. MPI still provides the best performance under some conditions. We present breakdowns of the execution times and measurements of hardware performance counters to explain the performance differences. Copyright © 2005 John Wiley & Sons, Ltd.     

Parallel strategies for 2D Discrete Wavelet Transform in shared memory systems and GPUs

In this work, we analyze the behavior of several parallel algorithms developed to compute the two-dimensional discrete wavelet transform using both OpenMP over a multicore platform and CUDA over a GPU. The proposed parallel algorithms are based on both regular filter-bank convolution and lifting transform with small implementations changes focused on both the memory requirements reduction and the complexity reduction. We compare our implementations against sequential CPU algorithms and other recently proposed algorithms like the SMDWT algorithm over different CPUs and the Wippig&Klauer algorithm over a GTX280 GPU. Finally, we analyze their behavior when algorithms are adapted to each architecture. Significant execution times improvements are achieved on both multicore platforms and GPUs. Depending on the multicore platform used, we achieve speed-ups of 1.9 and 3.4 using two and four processes, respectively, when compared to the sequential CPU algorithm, or we obtain speed-ups of 7.1 and 8.9 using eight and ten processes. Regarding GPUs, the GPU convolution algorithm using the GPU shared memory obtains speed-ups up to 20 when compared to the CPU sequential algorithm.     

Combining Data and Computation Distribution Directives for Hybrid Parallel Programming : A Transformation System

This paper describes dSTEP, a directive-based programming model for hybrid shared and distributed memory machines. The originality of our work is the definition and an implementation of a unified high-level programming model addressing both data and computation distributions, providing a particularly fine control of the computation. The goal is to improve the programmer productivity while providing good performances in terms of execution time and memory usage. We define a generic compilation scheme for computation mapping and communication generation. We implement the solution in a source-to-source compiler together with a runtime library. We provide a series of optimizations to improve the performance of the generated code, with a special focus on reducing the communications time. We evaluate our solution on several scientific kernels as well as on the more challenging NAS BT benchmark, and compare our results with the hand written Fortran MPI and UPC implementations. The results show first that our solution allows to make explicit the non trivial parallel execution of the NAS BT benchmark using the dSTEP directives. Second, the results show that our generated MPI+OpenMP BT program runs with a 83.35 speedup over the original NAS OpenMP C benchmark on a hybrid cluster composed of 64 quadricores (256 cores). Overall, our solution dramatically reduces the programming effort while providing good time execution and memory usage performances. This programming model is suitable for a large variety of machines as multi-core and accelerator clusters.     

dOpenCL: Towards uniform programming of distributed heterogeneous multi-/many-core systems

Modern computer systems become increasingly distributed and heterogeneous by comprising multi-core CPUs, GPUs, and other accelerators. Current programming approaches for such systems usually require the application developer to use a combination of several programming models (e.g., MPI with OpenCL or CUDA) in order to exploit the system’s full performance potential. In this paper, we present dOpenCL (distributed OpenCL)—a uniform approach to programming distributed heterogeneous systems with accelerators. dOpenCL allows the user to run unmodified existing OpenCL applications in a heterogeneous distributed environment. We describe the challenges of implementing the OpenCL programming model for distributed systems, as well as its extension for running multiple applications concurrently. Using several example applications, we compare the performance of dOpenCL with MPI + OpenCL and standard OpenCL implementations.     

Assessment of offload-based programming environments for hybrid CPU–MIC platforms in numerical modeling of solidification

Heterogeneous (or hybrid) computing platforms with Intel Xeon Phi accelerators offer potential advantages of energy efficient, massively parallel computing, while supporting parallel programming models familiar for users of multicore CPUs. However, realizing this potential for real-world applications still remains a challenging issue. The main goal of this paper is the suitability assessment of offload-based programming environments for porting a real-life scientific application to hybrid platforms with Intel KNC and KNL accelerators, assuming no significant modifications of the application code. The main criterion of this assessment is the application performance. The evaluated environments include: 1) Intel Offload coupled with OpenMP, 2) OpenMP 4.0 and 3) OpenMP 4.5 Accelerator Models, and 4) hStreams Library with OpenMP. A real-life application dedicated to the numerical modeling of alloy solidification is used as a testbed in the assessment. An experimental evaluation of the four versions of the application code for a platform with KNC coprocessors shows that excluding OpenMP 4.0, the rest of them are able to adapt to expansion of available resources, however, with different efficiency. While the shortest execution time is achieved for Intel Offload, the high-level abstractions of hStreams contribute considerably to making porting and tuning the application easier, with low performance overheads in comparison to the low-level Intel Offload extension. Benchmarking the application performance and scalability on a platform with multiple KNL processors, using the Offload over Fabric technology with Intel Offload and OpenMP 4.5, concludes the assessment.     

异构集群上的宏基因组聚类优化

宏基因组基因聚类是筛选致病基因的新型方法,其依赖于海量的测序数据、有效的聚类算法以及高效的计算机来实现。相关系数矩阵的计算是进行聚类前必须完成的操作,占总计算量的比重较大。以某基因库为例,包含1300个样本、每样本百万基因的数据,单线程运行需要27年。充分发挥多核CPU的潜力,利用GPU加速卡强大的计算能力,将程序扩展到多节点集群上运行,是重要而迫切的工作。在仔细分析算法的基础上,首先针对单CPU节点和单GPU卡做了高效实现,获得了接近理想的加速比;然后利用缓存优化进一步提升性能;最后使用负载均衡方法在MPI线程间分发计算任务,实现了良好的扩展。相比未优化的单线程程序,16节点CPU获得了238.8倍的加速,6 块GPU卡获得了263.8倍的加速。