高性能计算集群中的以太网技术

需积分: 1 82 浏览量更新于2024-09-22 收藏 363KB PDF 举报

"本文档主要探讨了在高性能计算集群（High Performance Computing Clusters，HPCC）中，以太网（Ethernet）的角色与应用。文档旨在澄清术语混淆，定义关键概念，并讨论与云网络（Cloud Networking）、交换机（Switches）相关的网络问题。" 在高性能计算领域，术语众多且有时会引起混淆。首先，我们需要明确几个基本概念： 1. HPC（High Performance Computing）：这涵盖了所有用于执行高需求任务的计算机系统，无论是高度优化的单一超级计算机还是由标准服务器组成的集群。 2. HPCC（High Performance Computing Cluster）：特指由多台计算机组成的集群，它们共同处理大型任务。HPCC同时是HPC的一个子集，也是一种特殊的超级计算机形式。 3. 集群（Cluster）：大致来说，集群是由一组同构的完整计算机系统组成，它们协同工作，将计算任务分解并并行处理，作为一个统一的计算资源。以太网在高性能计算集群中的作用至关重要，因为它提供了连接这些计算机节点的基础。以下是一些关键知识点： 1. **网络架构**：在集群中，以太网被广泛用作通信骨干，连接各个计算节点，允许它们快速交换数据。这可以是传统的共享介质以太网，也可以是更高级的如千兆以太网、万兆以太网，甚至更高带宽的40G或100G以太网。 2. **交换技术**：以太网交换机在HPCC中扮演核心角色，它们负责数据包的转发和路由，确保信息正确无误地发送到目标节点。高性能交换机通常具有低延迟、高带宽和可扩展性等特点，以满足大量并发通信的需求。 3. **协议与标准**：TCP/IP协议栈是集群通信的基础，但为了优化高性能计算，还发展了诸如InfiniBand、RoCE（RDMA over Converged Ethernet）等协议，它们提供更低的延迟和更高的吞吐量。特别是RoCE，它将远程直接内存访问（RDMA）技术集成到以太网中，进一步提高了性能。 4. **负载均衡与容错**：通过以太网连接的集群可以通过负载均衡技术分配任务，确保资源得到充分利用而不会过载。同时，通过冗余连接和故障切换机制，可以提高系统的可用性和可靠性。 5. **云计算集成**：随着云网络的发展，以太网成为连接数据中心内部和跨数据中心的关键技术。在云环境中，HPCC可以利用弹性云服务和按需扩展的资源，以支持大规模并行计算。 6. **软件定义网络（SDN）与网络功能虚拟化（NFV）**：这些新兴技术允许更灵活地管理和配置以太网网络，提高网络效率，降低运维成本。在HPCC中，SDN和NFV可以动态调整网络资源，以适应计算任务的变化。以太网在高性能计算集群中扮演着核心角色，连接并协调各个计算节点，支持高速数据传输和复杂的计算任务。随着技术的进步，以太网在性能、可扩展性和灵活性方面持续提升，以满足不断增长的高性能计算需求。

Ethernet in High Performance Computing Clusters

P AGE 1 OF 9

]

Terminology Confusion

Let us begin with a few words about the terminology. HPC,

HPCC, clusters, grids – this part of the computer world is

full of jargon and conflicting definitions. Before we talk

about networking issues, let’s put some definitions in place

and speak a common language. There are many strong

opinions in the computing community about defining the

taxonomy correctly – let’s use these as working definitions:

HPC – High Performance Computing

This term includes any computer constructed to complete

demanding tasks, whether it’s a highly tuned single super-

computer system or a cluster of off-the-shelf servers.

HPCC – High Performance Computing Cluster

This specifically refers to a cluster of computers assem-

bled to take on a larger task (it is also defined as High

Performance Computing and Communications). Any

HPCC is also an HPC, a cluster and a supercomputer.

Cluster

Loosely, a cluster is a group of homogeneous, whole

computer systems running in concert to divide and conquer

PAPER

White

a computing task and used as a unified computing resource.

These can include specially built high performance

computing nodes or off-the-shelf PCs or servers. An HPCC

is a special case of clustering – a very large cluster.

Supercomputer

This term is used widely and defined loosely. Today the

definition of a supercomputer is a computing system fast

enough to put it among the world’s fastest, whether it’s a

single highly tuned computer or a cluster of off-the-shelf

servers.

Grid

This term means many things to many people. In some

cases, it’s a group of clusters spread across a network, as

in the NSF-funded TeraGrid project. The TeraGrid project

is extending the frontiers of supercomputing, harnessing

and using distributed computers, data storage systems,

networks, and other resources as if they were a single

massive system, regardless of physical location, and

creating "virtual supercomputers."

Abstract

Clustering has emerged as a cost-effective architecture for building high performance computing facilities. To keep

pace with this trend, the requirements for data center network equipment have evolved rapidly in an effort to keep

pace with the four types of connections typically required in a cluster. These connections are I/O, out-of-band manage-

ment, storage, and interconnect, and each has different design parameters and requirements. Traditionally, Ethernet was

required in every cluster – providing some or all of the connections, depending on the cluster architecture. Recently,

innovations in Ethernet make it an ideal connection for all of these connections. This paper examines the issues and

tradeoffs in different technologies used for these connections.

Introduction

Ethernet continues to grow as the dominant and de facto standard for networking. It’s shipped on every PC and

server-class system, and is present in every broadband household. It has taken over corporate LANs to the point where

it is challenging to find alternate LAN technologies (remember Token Ring? FDDI? Arcnet?). Today, Ethernet is rapidly

solving new problems in MAN, WAN, and in wireless applications.

Indeed, R&D spending on Ethernet outpaces spending on all other networking technologies combined, resulting in

continuously improving performance, higher speeds, new capabilities, rapidly falling price points, and rapid evolution

to meet new requirements.

In this paper we’ll look at the issues and get some insights into cluster computing while asking the question: When

is Ethernet the right technology for building clusters? We’ll look at the four main categories of communications needs

for clusters – I/O, out-of-band management, storage, and interconnects – and the pros and cons of the different

technologies available for these different categories.

下载后可阅读完整内容，剩余8页未读，立即下载

b123enben

粉丝: 0
资源: 19

高性能计算集群中的以太网技术

example_wp_log_peyton_manning.csv

Ajax-WP_AJAX.zip

wordpress 数据表wp_posts，wp_postmeta，wp_terms，wp_term_relationships，wp_term_taxonomy，wp_links，这些表字段，哪些是无用的，在注释里标注出来。

WP_Route：一个简单的类，用于将复杂的路由绑定到函数方法或WP_AJAX操作

wp-combine-queries：WordPress：通过WP_Query类的自定义'combined_query'属性将多个WP_Query（）查询合并为一个查询

wp_area_traj_3=[] for i in wp_area_traj: a=np.array(i) wp_area_traj_3.append(a)表示什么意思

arkainte_wp_onegen_sql_源码

wp_complex_bone_theme

最新资源