<?xml version="1.0" encoding="UTF-8"?>

<rdf:RDF 
  xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
  xmlns="http://purl.org/rss/1.0/"
  xmlns:dc="http://purl.org/dc/elements/1.1/"
  xmlns:syn="http://purl.org/rss/1.0/modules/syndication/"
>

  <channel rdf:about="http://lwn.net/headlines/254445/">
    <title>LWN: Comments on "Memory part 4: NUMA support"</title>
    <link>http://lwn.net/Articles/254445/</link>
    <description>
This is a special feed containing comments posted
to the individual LWN article titled &quot;Memory part 4: NUMA support&quot;.

    </description>

    <syn:updatePeriod>hourly</syn:updatePeriod>
    <syn:updateFrequency>2</syn:updateFrequency>
    <items>
      <rdf:Seq>
	<rdf:li resource="http://lwn.net/Articles/264585/rss" />
	<rdf:li resource="http://lwn.net/Articles/256207/rss" />
	<rdf:li resource="http://lwn.net/Articles/255863/rss" />
	<rdf:li resource="http://lwn.net/Articles/255293/rss" />
	<rdf:li resource="http://lwn.net/Articles/255112/rss" />
	<rdf:li resource="http://lwn.net/Articles/254969/rss" />
      
      </rdf:Seq>
    </items>

  </channel>
    <item rdf:about="http://lwn.net/Articles/264585/rss">
      <title>Memory part 4: NUMA support</title>
      <link>http://lwn.net/Articles/264585/rss</link>
      <dc:date>2008-01-10T14:27:44+00:00</dc:date>
      <dc:creator>rengolin</dc:creator>
      <description>
      &lt;div class=&quot;FormattedComment&quot;&gt;&lt;pre&gt;
&lt;font class=&quot;QuotedText&quot;&gt;&amp;gt; a radius 1 arrangement can actually include 3 CPUs, think a triangle, every machine is at
most one hop from the memory&lt;/font&gt;

Actually it's N+1 being N the number of dimensions you build your computer. 

In the three-dimensional reality a tetrahedron (4CPUs) is radius 1, in the two-dimensional a
triangle and so on, therefore we can increase the number of CPUs in radius 1 by building
computers in higher dimensions! ;)
&lt;/pre&gt;&lt;/div&gt;

      
      </description>
    </item>
    <item rdf:about="http://lwn.net/Articles/256207/rss">
      <title>Memory part 4: NUMA support</title>
      <link>http://lwn.net/Articles/256207/rss</link>
      <dc:date>2007-10-27T06:51:09+00:00</dc:date>
      <dc:creator>dlang</dc:creator>
      <description>
      it's also possible to have a radius-3 arrangement with only 3 links
&lt;br&gt;&lt;pre&gt;
1--2--
|  | |
3--4 |
|  | |
5--6 |
|  | |
7--8--
&lt;/pre&gt;&lt;br&gt;
this leaves CPU's 1 and 7 with an extra link available for connections to a southbridge.
      
      </description>
    </item>
    <item rdf:about="http://lwn.net/Articles/255863/rss">
      <title>Memory part 4: NUMA support</title>
      <link>http://lwn.net/Articles/255863/rss</link>
      <dc:date>2007-10-25T12:47:08+00:00</dc:date>
      <dc:creator>joib</dc:creator>
      <description>
      &lt;i&gt;
the Opteron already has versions that have four HT links, they are the 800/8000 series chips
targeted at 8 cpu systems
&lt;/i&gt;
&lt;p&gt;
That is not correct. All current Opterons (including the 800(0) series) have 3 HT links. The difference is that in the 800(0) series all 3 HT links are cache coherent, whereas there is only one cache coherent link on the 200(0) series and none on the 100(0) series (and athlon64).
&lt;p&gt;
That being said, the recently released quad core Opterons do in principle support 4 HT links, but so far they are using the existing socket so they are limited to 3 links until the next generation socket is introduced.
      
      </description>
    </item>
    <item rdf:about="http://lwn.net/Articles/255293/rss">
      <title>Memory part 4: NUMA support</title>
      <link>http://lwn.net/Articles/255293/rss</link>
      <dc:date>2007-10-22T00:35:12+00:00</dc:date>
      <dc:creator>dlang</dc:creator>
      <description>
      &lt;pre class=&quot;FormattedComment&quot;&gt;
several comments

a radius 1 arrangement can actually include 3 CPUs, think a triangle, every machine is at most
one hop from the memory

the Opteron already has versions that have four HT links, they are the 800/8000 series chips
targeted at 8 cpu systems
&lt;/pre&gt;

      
      </description>
    </item>
    <item rdf:about="http://lwn.net/Articles/255112/rss">
      <title>Memory part 4: NUMA support</title>
      <link>http://lwn.net/Articles/255112/rss</link>
      <dc:date>2007-10-19T04:47:10+00:00</dc:date>
      <dc:creator>sweikart</dc:creator>
      <description>
      &lt;pre class=&quot;FormattedComment&quot;&gt;
Intel is working on a point-to-point processor interconnect called QuickPath, which is a
competitor to AMD's HyperTransport.

AMD will support HyperTransport 3.0 with its next-generation sockets (AM2+ and F+).
HyperTransport 3.0 lets you split the links in half, so you can have 8 links total (and HT3.0
links can run twice as fast as HT2.0 links, so the half-width links should have good
throughput).  I assume this will enable a fully-connected (diameter of 1) eight-socket server.

&lt;/pre&gt;

      
      </description>
    </item>
    <item rdf:about="http://lwn.net/Articles/254969/rss">
      <title>Memory part 4: NUMA support</title>
      <link>http://lwn.net/Articles/254969/rss</link>
      <dc:date>2007-10-18T13:33:18+00:00</dc:date>
      <dc:creator>dany</dc:creator>
      <description>
      &lt;pre class=&quot;FormattedComment&quot;&gt;
Thanks for another very informative article!
I am wondering, in relevant opteron CPU measured, how much slower are read/write operations on
remote memory 2-hops away (figure 5.4 is about 1 hop away remote memory)?
&lt;/pre&gt;

      
      </description>
    </item>
</rdf:RDF>

