CINXE.COM

<?xml version="1.0" encoding="UTF-8" standalone="yes"?> <!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"><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8" /><title>LKML: Paul Jackson: Re: Industry db benchmark result on recent 2.6 kernels</title><link href="/css/message.css" rel="stylesheet" type="text/css" /><link href="/css/wrap.css" rel="alternate stylesheet" type="text/css" title="wrap" /><link href="/css/nowrap.css" rel="stylesheet" type="text/css" title="nowrap" /><link href="/favicon.ico" rel="shortcut icon" /><script src="/js/simple-calendar.js" type="text/javascript"></script><script src="/js/styleswitcher.js" type="text/javascript"></script><link rel="alternate" type="application/rss+xml" title="lkml.org : last 100 messages" href="/rss.php" /><link rel="alternate" type="application/rss+xml" title="lkml.org : last messages by Paul Jackson" href="/groupie.php?aid=5266" /><script> var _paq = window._paq = window._paq || []; /* tracker methods like "setCustomDimension" should be called before "trackPageView" */ _paq.push(["setDoNotTrack", true]); _paq.push(["disableCookies"]); _paq.push(['trackPageView']); _paq.push(['enableLinkTracking']); (function() { var u="//m.lkml.org/"; _paq.push(['setTrackerUrl', u+'matomo.php']); _paq.push(['setSiteId', '1']); var d=document, g=d.createElement('script'), s=d.getElementsByTagName('script')[0]; g.async=true; g.src=u+'matomo.js'; s.parentNode.insertBefore(g,s); })(); </script></head><body onload="es.jasper.simpleCalendar.init();" itemscope="itemscope" itemtype="http://schema.org/BlogPosting"><table border="0" cellpadding="0" cellspacing="0"><tr><td width="180" align="center"><a href="/"><img style="border:0;width:135px;height:32px" src="/images/toprowlk.gif" alt="lkml.org" /></a></td><td width="32">聽</td><td class="nb"><div><a class="nb" href="/lkml"> [lkml]</a> 聽 <a class="nb" href="/lkml/2005"> [2005]</a> 聽 <a class="nb" href="/lkml/2005/4"> [Apr]</a> 聽 <a class="nb" href="/lkml/2005/4/1"> [1]</a> 聽 <a class="nb" href="/lkml/last100"> [last100]</a> 聽 <a href="/rss.php"><img src="/images/rss-or.gif" border="0" alt="RSS Feed" /></a></div><div>Views: <a href="#" class="nowrap" onclick="setActiveStyleSheet('wrap');return false;">[wrap]</a><a href="#" class="wrap" onclick="setActiveStyleSheet('nowrap');return false;">[no wrap]</a> 聽 <a class="nb" href="/lkml/mheaders/2005/4/1/16" onclick="this.href='/lkml/headers'+'/2005/4/1/16';">[headers]</a>聽 <a href="/lkml/bounce/2005/4/1/16">[forward]</a>聽 </div></td><td width="32">聽</td></tr><tr><td valign="top"><div class="es-jasper-simpleCalendar" baseurl="/lkml/"></div><div class="threadlist">Messages in this thread</div><ul class="threadlist"><li class="root"><a href="/lkml/2005/3/28/161">First message in thread</a></li><li><a href="/lkml/2005/3/31/266">"Chen, Kenneth W"</a><ul><li><a href="/lkml/2005/3/31/295">Nick Piggin</a><ul><li class="origin"><a href="/lkml/2005/4/1/20">Paul Jackson</a><ul><li><a href="/lkml/2005/4/1/20">Nick Piggin</a><ul><li><a href="/lkml/2005/4/1/29">Paul Jackson</a></li></ul></li><li><a href="/lkml/2005/4/1/24">Ingo Molnar</a><ul><li><a href="/lkml/2005/4/1/264">"Chen, Kenneth W"</a></li></ul></li><li><a href="/lkml/2005/4/1/28">Ingo Molnar</a><ul><li><a href="/lkml/2005/4/1/59">Paul Jackson</a></li></ul></li></ul></li></ul></li><li><a href="/lkml/2005/3/31/349">Ingo Molnar</a><ul><li><a href="/lkml/2005/4/1/4">"Chen, Kenneth W"</a></li></ul></li></ul></li></ul></td><td width="32" rowspan="2" class="c" valign="top"><img src="/images/icornerl.gif" width="32" height="32" alt="/" /></td><td class="c" rowspan="2" valign="top" style="padding-top: 1em"><table><tr><td><table><tr><td class="lp">Date</td><td class="rp" itemprop="datePublished">Thu, 31 Mar 2005 22:05:26 -0800</td></tr><tr><td class="lp">From</td><td class="rp" itemprop="author">Paul Jackson <></td></tr><tr><td class="lp">Subject</td><td class="rp" itemprop="name">Re: Industry db benchmark result on recent 2.6 kernels</td></tr></table></td><td></td></tr></table><pre itemprop="articleBody">Nick wrote:<br />> Ingo had a cool patch to estimate dirty => dirty cacheline transfer latency<br />> ... Unfortunately ... and it is an O(cpus^2) operation.<br /><br />Yes - a cool patch.<br /><br />If we had an arch-specific bit of code, that for any two cpus, could<br />give a 'pseudo-distance' between them, where the only real requirements<br />were that (1) if two pairs of cpus had the same pseudo-distance, then<br />that meant they had the same size, layout, kind and speed of bus amd<br />cache hardware between them (*), and (2) it was cheap - hardly more than<br />a few lines of code and a subroutine call to obtain, then Ingo's code<br />could be:<br /><br /> for each cpu c1:<br /> for each cpu c2:<br /> psdist = pseudo_distance(c1, c2)<br /> if I've seen psdist before, use the latency computed for that psdist<br /> else compute a real latency number and remember it for that psdist<br /><br />A generic form of pseudo_distance, which would work for all normal<br />sized systems, would be:<br /><br />int pseudo_distance(int c1, int c2)<br />{<br /> static int x;<br /> return x++;<br />}<br /><br />Then us poor slobs with big honkin numa iron could code up a real<br />pseudo_distance() routine, to avoid the actual pain of doing real work<br />for cpus^2 iterations for large cpu counts.<br /><br />Our big boxes have regular geometries with much symmetry, so would<br />provide significant opportunity to exploit equal pseudo-distances.<br /><br />And I would imagine that costs of K * NCPU * NCPU are tolerable in this<br />estimation routine. for sufficiently small K, and existing values of<br />NCPU.<br /><br />(*) That is, if pseudo_distance(c1, c2) == pseudo_distance(d1, d2), then<br /> this meant that however c1 and c2 were connected to each other in the<br /> system (intervening buses and caches and such), cpus d1 and d2 were<br /> connected the same way, so could be presumed to have the same latency,<br /> close enough.<br /><br />-- <br /> I won't rest till it's the best ...<br /> Programmer, Linux Scalability<br /> Paul Jackson <pj@engr.sgi.com> 1.650.933.1373, 1.925.600.0401<br />-<br />To unsubscribe from this list: send the line "unsubscribe linux-kernel" in<br />the body of a message to majordomo@vger.kernel.org<br />More majordomo info at <a href="http://vger.kernel.org/majordomo-info.html">http://vger.kernel.org/majordomo-info.html</a><br />Please read the FAQ at <a href="http://www.tux.org/lkml/">http://www.tux.org/lkml/</a><br /><br /></pre></td><td width="32" rowspan="2" class="c" valign="top"><img src="/images/icornerr.gif" width="32" height="32" alt="\" /></td></tr><tr><td align="right" valign="bottom"> 聽 </td></tr><tr><td align="right" valign="bottom">聽</td><td class="c" valign="bottom" style="padding-bottom: 0px"><img src="/images/bcornerl.gif" width="32" height="32" alt="\" /></td><td class="c">聽</td><td class="c" valign="bottom" style="padding-bottom: 0px"><img src="/images/bcornerr.gif" width="32" height="32" alt="/" /></td></tr><tr><td align="right" valign="top" colspan="2"> 聽 </td><td class="lm">Last update: 2005-04-06 13:31 聽聽 [from the cache]<br />漏2003-2020 <a href="http://blog.jasper.es/"><span itemprop="editor">Jasper Spaans</span></a>|hosted at <a href="https://www.digitalocean.com/?refcode=9a8e99d24cf9">Digital Ocean</a> and my Meterkast|<a href="http://blog.jasper.es/categories.html#lkml-ref">Read the blog</a></td><td>聽</td></tr></table><script language="javascript" src="/js/styleswitcher.js" type="text/javascript"></script></body></html>