神威太湖之光上NAMD分子动力学软件的移植与优化

需积分: 13 70 浏览量更新于2024-09-05 收藏 774KB PDF 举报

"Porting and Optimizing NAMD on Sunway TaihuLight Supercomputer" 本文主要讨论了将分子动力学模拟软件NAMD（Nanoscale Molecular Dynamics）移植和优化到神威太湖之光超级计算机上的过程。NAMD是一款开源的分子动力学模拟工具，能够在大规模并行计算系统中模拟数百万个原子的宏观分子。它通过原子分区和力分解来实现负载均衡的并行计算。神威太湖之光超级计算机具有极高的理论峰值性能，达到125.4 Pﬂop/s，并且其LINPACK效率不低于70%，拥有1000万个处理器核心。这种计算能力为高性能科学计算提供了强大支持。针对神威太湖之光的独特架构，作者们进行了深入的性能分析和优化工作。结果显示，优化后的NAMD在神威太湖之光上的整体性能提升了将近30倍，单核性能比Intel Xeon E5-2650V2提升了两倍以上。这一优化突破意味着在大规模并行计算时，能够更有效地利用硬件资源。论文中还提到，移植后的NAMD理论扩展性达到了约180万个核心，这在超算应用领域是一个显著的成就。这表明NAMD在神威太湖之光上可以实现高效、线性的扩展，意味着随着核心数量的增加，软件性能也能保持同步增长，这对于处理大规模的分子动力学问题至关重要。此外，文章可能还涵盖了以下知识点： 1. **众核优化**：神威太湖之光采用众核架构，优化NAMD时需要考虑如何有效地分配任务到各个核心，确保并行计算的效率和负载均衡。 2. **并行算法设计**：NAMD的并行策略可能涉及到数据分片、通信协议和同步机制的调整，以适应神威太湖之光的通讯网络架构。 3. **性能监控与调优**：通过性能分析工具，对计算过程中的内存访问、计算效率、通信开销等关键指标进行监控，以便找出性能瓶颈并进行针对性优化。 4. **能耗效率**：在超级计算机领域，不仅关注峰值性能，也注重能效比。优化NAMD在神威太湖之光上的运行可能涉及降低能耗和提高计算效率。 5. **跨平台兼容性**：NAMD作为开源软件，其移植工作涉及到代码的跨平台适配，确保在不同架构的系统上都能稳定运行。 6. **科学计算应用**：NAMD在生物物理学、药物研发、材料科学等领域有广泛应用，优化后的版本将为这些领域的研究提供更快更准确的计算支持。这篇论文提供了将高性能科学计算软件NAMD优化到神威太湖之光超级计算机的实践经验，对于理解如何在特定超级计算机平台上进行高效并行计算和优化具有重要的参考价值。同时，也为其他类似的高性能应用提供了移植和优化的思路。

Porting and Optimizing NAMD on Sunway TaihuLight Supercomputer

Wenjun Yao, Weikeng Chen, Zhichao Su, Junshi Chen, Chenzhi Liao and Hong An

Department of Computer Science and Technology, University of Science and Technology of China

Email: {yaowj,cwk32,suzc,chenjs,czliao}@mail.ustc.edu.cn, han@ustc.edu.cn

Abstract—NAMD (Nanoscale Molecular Dynamics) an open-

source molecular dynamics simulation software, which can

simulate millions of atoms macromolecules in massively parallel

computer system. The NAMD uses atom partition and force

division to utilize the parallel computing with load balance

control. In this paper, we focus on the porting and optimizing

of NAMD to one of top supercomputers, Sunway TaihuLight.

Sunway TaihuLight is a supercomputer with a theoretical

peak performance of 125.4 Pﬂop/s and LINPACK efﬁciency

no less than 70%, and has 10 million cores. Performance

analysis shows that the total performance is increased nearly

30 times, and the single-core performance is twice more than

that of Intel Xeon E5-2650 V2. The theoretical scalability of our

porting reached about 1.8 million core, which is a breakthrough

compared with previous experiments on NAMD.

Keywords-Molecular dynamics simulation; Heterogeneous

many-core processor;

I. INTRODUCTION

NAMD (Nanoscale Molecular Dynamics) [1], [2] is a

molecular dynamics simulation software which supports the

simulation of millions of atoms in large-scale parallel com-

puter system. It uses forceﬁelds (like Amber, CHARMM,

and Dreiding) for numerical solution of motion equations

to simulate the motion of atoms. It can simulate molecular

dynamics in different scales and is widely used in the

research of computational chemistry [3], [4].

Sunway TaihuLight supercomputer [5] is the top super-

computer developed by National Research Center of Parallel

Computer Engineering & Technology (NRCPC) with peak

performance of 125.4 Pﬂop/s. The supercomputer under

linpack benchmark achieves peak ﬂops/W at 6 Gﬂops/W and

efﬁciency at 74.15%. The Sunway supercomputer provides

two types of nodes in the system. The ﬁrst type is the high-

performance computing node with the ALPHA architecture,

which is specially used for computation. The second type is

the management node under the commercial Intel platform,

which is used for system management, software develop-

ment, and job submission. In the computer node, each CPU

has four core groups (CGs), and in each one core group

there are one management processing element (MPE) as

main cores and 64 computing processing element (CPEs) as

slave cores that shares the same memory space of that core

group. The computing nodes in Sunway supercomputer are

managed by those Intel platform management nodes. When

developers want job submission, they log in to the remote

management nodes which are with Intel platform, use the

cross compilation tools to make application programs, and

submit the job using bsub command.

In this paper, we port NAMD to the Sunway TaihuLight

supercomputer and optimize for better efﬁciency in simu-

lation. We need to port three third-party libraries that are

used in NAMD before porting the main program: TCL,

FFTW, and CHARM++ [6], [7], which is ported by cross

compilation. Here the CHARM++ is the parallel program-

ming framework supporting NAMD that manages the task

scheduling and the load balancing. We modify the NAMD

to make use of the instructions of Sunway CPU to boost the

main computation.

II. SUNWAY CPU ARCHITECTURE

The Sunway supercomputer uses the processors produced

by Shanghai High Performance IC Design Center. In the

aspect of software, it has a customized x64 Linux operating

system with the compiler for basic languages like C/C++,

Fortran, and some parallel programming languages like MPI,

OpenMP, and OpenACC. The Sunway also provides a high-

performance thread library for the ﬁne-grained management

of parallel computing. The Sunway processors work with

the frequency of 1.5GHz.

Figure 1. Sunway Core Group and Sunway Processor

Each processor has four core groups (CGs), and each core

group has one management processing element (MPE) and

8×8 array of computing processing elements (CPEs). Each

one processor has a memory controllers (MC) for memory

access. These components on the CPU are interconnected.

For four core groups in one processor, they are connected

by network on chip (NoC) with private memory space. For

elements in a core group, MPE and CPEs are connected

by the memory controllers (MC) and the memory space.

All processors are interconnected through a system interface

(SI). The structure of a core group and a Sunway CPU is

showed in Fig. 1.

下载后可阅读完整内容，剩余7页未读，立即下载

小子他亲爹

粉丝: 0
资源: 5

神威太湖之光上NAMD分子动力学软件的移植与优化

[逆向工程资源].Syngress.Sockets.Shellcode.Porting.And.Coding.

captronic-porting-linux-on-arm.pdf

蓝牙协议栈到mcu的移植 最新英文文献?

porting layer时间什么

..\Middlewares\LVGL\GUI\lvgl\examples\porting\lv_port_disp_template.c(144): error: #171: invalid type conversion

ERROR: Could not find a version that satisfies the requirement PyXML (from versions: none) ERROR: No matching distribution found for PyXML

porting security

stm32移植freertos

蓝牙协议栈到mcu的移植 最新文献?

最新资源

蓝牙协议栈到mcu的移植最新英文文献?

蓝牙协议栈到mcu的移植最新文献?