📄 12.html

📁 国外MPI教材
💻 HTML
字号:
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"    "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"><html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en" lang="en"><head>	<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1" />	<style type="text/css">	body { font-family: Verdana, Arial, Helvetica, sans-serif;}	a.at-term {	font-style: italic; }	</style>	<title>Examples of Distributed Shared Memory Systems</title>	<meta name="Generator" content="ATutor">	<meta name="Keywords" content=""></head><body> <h3> NUMA Systems </h3>

<h4> HP Exemplar and Superdome </h4>

<p> The first widely available NUMA system was the <a href="http://www.hp.com/" target="_blank">HP</a> 
(formerly Convex) Exemplar. The Exemplar could have up to 64 processors in a single system. The system was composed of "hypernodes" of either 4 or 8 PA-RISC 7100 processors connected to a local memory bank via a crossbar switch. These hypernodes were connected together in a ring topology using a <a href="http://sunrise.scu.edu/WhatIsSCI.html">Scalable 
Coherent Interconnect (SCI)</a> network. </p>

<p> A follow-on to the Exemplar, called the <a href="http://www.hp.com/products1/unixservers/highend/superdome/index.html"> 
Superdome</a>, is currently available from HP. Like the Exemplar, the Superdome can have up to 64 processors in a single system. The major difference between the Exemplar and Superdome architectures is that the Superdome uses a hierarchy of crossbars rather than an SCI ring as its topology. The latency cost of a remote memory access on the Superdome is approximately 1.6 times that of a local memory access. </p>

<h3> SGI Origin 2000 and 3000 </h3>

<p> Currently (June, 2001) the most widely used NUMA system in HPC circles is the <a href="http://www.sgi.com/">SGI</a> <a href="http://www.sgi.com/origin/2000/">Origin 2000</a> (O2k). A standard O2k configuration could have up to 128 processors in a single system, although custom installations of up to 512 processors in a single system exist. The O2k is composed of "nodes" of two MIPS R12k processors connected to a local memory bank via a crossbar switch. The topology of the system is a hypercube of routers, where each router can have two nodes directly connected to it. In configurations larger than 64 processors, the hypercube router topology is replaced by hypercubes of routers connected through metarouters. The latency cost of a remote memory access is at most 3 times that of a local memory access. </p>

<p> SGI has a follow-on to the O2k called the <a href="http://www.sgi.com/origin/3000/">Origin 3000</a> (O3k). The O3k has 4 processors per "C brick" node instead of 2, and "R brick" routers can connect to 4 C bricks instead of 2. The bandwidth to local and hemote memory is also improved. O3k systems can have up to 512 processors in a standard configuration, and custom configurations of 1024 processors are expected. The ratio of remote to local memory latency on the O3k is similar to that of the O2k. </p>

<h3> Compaq AlphaServer GS320 </h3>

<p> A relative newcomer to the NUMA market is the <a href="http://www.compaq.com/">Compaq</a> <a href="http://www.compaq.com/alphaserver/gs320/index.html">AlphaServer GS320</a>. The GS320 can have up to 32 processors in a single system. The system is composed of blocks of 4 Alpha EV6 processors connected to a local memory bank via a crossbar switch. These blocks are connected together with a central crossbar switch, similar to that of the HP Superdome. The ratio of remote to local memory latency in the GS320 is at most 3. </p>

<h3> Clusters of SMPs </h3>

<h4> IBM SP </h4>

<p> The <a href="http://www.ibm.com/servers/eserver/pseries/hardware/largescale/sp.html">IBM 
SP</a> is possibly the longest lived commercial parallel computer family in existence. SPs are clusters of RS/6000 nodes connected by a proprietary, hierarchically 
switched network. Early SPs were made up of uniprocessor nodes, but most current SPs have nodes with 4, 8, or 16 Power3 processors each. SPs have been scaled to very large node and processor counts, the largest of which are the <a href="http://www.llnl.gov/asci/platforms/white/">ASCI 
White</a> (512 nodes with 16 processors each) and <a href="http://www.llnl.gov/asci/platforms/bluepac/">ASCI 
Blue Pacific</a> (1800 nodes with 4 processors each). </p>

<h4> Compaq AlphaServer SC </h4>

<p> A comparatively new commercial parallel computer is the Compaq <a href="http://www.compaq.com/alphaserver/sc/">AlphaServer 
SC</a>. An AlphaServer SC is a cluster of up to 128 4-processor <a href="http://www.compaq.com/alphaserver/es40/">ES40</a> 
or ES45 nodes connected by a high speed interconnect from <a href="http://www.quadrics.com/">Quadrics</a>. 
Compaq plans to scale this to more than 600 nodes for the second phase of the <a href="http://www.psc.edu/machines/tcs/">PSC Terascale Computing system</a>. </p>

<h4> Beowulf Clusters </h4>

<p> Many of the larger "Beowulf" clusters of Intel- and Alpha-based systems running Linux are clusters of SMPs. Examples of this include: 
<ul>
  <li> <a  href="http://www.ahpcc.unm.edu/Computing/AHPCC/Machines/loslobos.html">Los 
Lobos</a> (256 2-way Pentium III nodes with <a
href="http://www.myri.com/myrinet/">Myrinet</a>) and <a
href="http://www.ahpcc.unm.edu/Computing/AHPCC/Machines/roadrunner.html"> 
Roadrunner</a> (32 2-way Penitum II nodes with Myrinet) at <a href="http://www.ahpcc.unm.edu/">Albuquerque High Performance Computing Center</a>. </li>
  <li> <a href="http://www-unix.mcs.anl.gov/chiba/index.html">Chiba City</a> (314 2-way Pentium III nodes with Myrinet) at <a href="http://www.anl.gov/">Argonne National Laboratory</a>. </li>
  <li> <a href="http://www.acl.lanl.gov/cluster/">The Little Blue Penguin Cluster</a> (64 2-way Pentium II nodes with Myrinet) at <a href="http://www.lanl.gov/">Los Alamos National Laboratory</a>. 
  <li> <a href="http://oscinfo.osc.edu/hardware/index.html#beo">The Brain</a> (33 4-way Pentium III nodes with Myrinet) and Wakko (73 2-way Itanium nodes with Myrinet) at <a href="http://www.osc.edu/">Ohio Supercomputer Center</a>. 
  </li>
</ul>
<p></p></body></html>
⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -