ترغب بنشر مسار تعليمي؟ اضغط هنا

Accelerating Concurrent Heap on GPUs

91   0   0.0 ( 0 )
 نشر من قبل Yanhao Chen
 تاريخ النشر 2019
  مجال البحث الهندسة المعلوماتية
والبحث باللغة English
 تأليف Yanhao Chen




اسأل ChatGPT حول البحث

Priority queue, often implemented as a heap, is an abstract data type that has been used in many well-known applications like Dijkstras shortest path algorithm, Prims minimum spanning tree, Huffman encoding, and the branch-and-bound algorithm. However, it is challenging to exploit the parallelism of the heap on GPUs since the control divergence and memory irregularity must be taken into account. In this paper, we present a parallel generalized heap model that works effectively on GPUs. We also prove the linearizability of our generalized heap model which enables us to reason about the expected results. We evaluate our concurrent heap thoroughly and show a maximum 19.49X speedup compared to the sequential CPU implementation and 2.11X speedup compared with the existing GPU implementation. We also apply our heap to single source shortest path with up to 1.23X speedup and 0/1 knapsack problem with up to 12.19X speedup.



قيم البحث

اقرأ أيضاً

Stencil computations are widely used in HPC applications. Today, many HPC platforms use GPUs as accelerators. As a result, understanding how to perform stencil computations fast on GPUs is important. While implementation strategies for low-order sten cils on GPUs have been well-studied in the literature, not all of proposed enhancements work well for high-order stencils, such as those used for seismic modeling. Furthermore, coping with boundary conditions often requires different computational logic, which complicates efficient exploitation of the thread-level parallelism on GPUs. In this paper, we study high-order stencils and their unique characteristics on GPUs. We manually crafted a collection of implementations of a 25-point seismic modeling stencil in CUDA and related boundary conditions. We evaluate their code shapes, memory hierarchy usage, data-fetching patterns, and other performance attributes. We conducted an empirical evaluation of these stencils using several mature and emerging tools and discuss our quantitative findings. Among our implementations, we achieve twice the performance of a proprietary code developed in C and mapped to GPUs using OpenACC. Additionally, several of our implementations have excellent performance portability.
Rapid growth in scientific data and a widening gap between computational speed and I/O bandwidth make it increasingly infeasible to store and share all data produced by scientific simulations. Instead, we need methods for reducing data volumes: ideal ly, methods that can scale data volumes adaptively so as to enable negotiation of performance and fidelity tradeoffs in different situations. Multigrid-based hierarchical data representations hold promise as a solution to this problem, allowing for flexible conversion between different fidelities so that, for example, data can be created at high fidelity and then transferred or stored at lower fidelity via logically simple and mathematically sound operations. However, the effective use of such representations has been hindered until now by the relatively high costs of creating, accessing, reducing, and otherwise operating on such representations. We describe here highly optimized data refactoring kernels for GPU accelerators that enable efficient creation and manipulation of data in multigrid-based hierarchical forms. We demonstrate that our optimized design can achieve up to 250 TB/s aggregated data refactoring throughput -- 83% of theoretical peak -- on 1024 nodes of the Summit supercomputer. We showcase our optimized design by applying it to a large-scale scientific visualization workflow and the MGARD lossy compression software.
135 - Frank Winter 2011
Extensions to the C++ implementation of the QCD Data Parallel Interface are provided enabling acceleration of expression evaluation on NVIDIA GPUs. Single expressions are off-loaded to the device memory and execution domain leveraging the Portable Ex pression Template Engine and using Just-in-Time compilation techniques. Memory management is automated by a software implementation of a cache controlling the GPUs memory. Interoperability with existing Krylov space solvers is demonstrated and special attention is paid on Chroma readiness. Non-kernel routines in lattice QCD calculations typically not subject of hand-tuned optimisations are accelerated which can reduce the effects otherwise suffered from Amdahls Law.
To accelerate the solution of large eigenvalue problems arising from many-body calculations in nuclear physics on distributed-memory parallel systems equipped with general-purpose Graphic Processing Units (GPUs), we modified a previously developed hy brid MPI/OpenMP implementation of an eigensolver written in FORTRAN 90 by using an OpenACC directives based programming model. Such an approach requires making minimal changes to the original code and enables a smooth migration of large-scale nuclear structure simulations from a distributed-memory many-core CPU system to a distributed GPU system. However, in order to make the OpenACC based eigensolver run efficiently on GPUs, we need to take into account the architectural differences between a many-core CPU and a GPU device. Consequently, the optimal way to insert OpenACC directives may be different from the original way of inserting OpenMP directives. We point out these differences in the implementation of sparse matrix-matrix multiplications (SpMM), which constitutes the main cost of the eigensolver, as well as other differences in the preconditioning step and dense linear algebra operations. We compare the performance of the OpenACC based implementation executed on multiple GPUs with the performance on distributed-memory many-core CPUs, and demonstrate significant speedup achieved on GPUs compared to the on-node performance of a many-core CPU. We also show that the overall performance improvement of the eigensolver on multiple GPUs is more modest due to the communication overhead among different MPI ranks.
128 - Frank Winter 2011
Graphic Processing Units (GPUs) are getting increasingly important as target architectures in scientific High Performance Computing (HPC). NVIDIA established CUDA as a parallel computing architecture controlling and making use of the compute power of GPUs. CUDA provides sufficient support for C++ language elements to enable the Expression Template (ET) technique in the device memory domain. QDP++ is a C++ vector class library suited for quantum field theory which provides vector data types and expressions and forms the basis of the lattice QCD software suite Chroma. In this work accelerating QDP++ expression evaluation to a GPU was successfully implemented leveraging the ET technique and using Just-In-Time (JIT) compilation. The Portable Expression Template Engine (PETE) and the C API for CUDA kernel arguments were used to build the bridge between host and device memory domains. This provides the possibility to accelerate Chroma routines to a GPU which are typically not subject to special optimisation. As an application example a smearing routine was accelerated to execute on a GPU. A significant speed-up compared to normal CPU execution could be measured.
التعليقات
جاري جلب التعليقات جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
mircosoft-partner

هل ترغب بارسال اشعارات عن اخر التحديثات في شمرا-اكاديميا