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: [PATCH] -mm tree: broken "dynamic sched domains" and "migration cost"</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/2005/12"> [Dec]</a> 聽 <a class="nb" href="/lkml/2005/12/9"> [9]</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/2005/12/9/255" onclick="this.href='/lkml/headers'+'/2005/12/9/255';">[headers]</a>聽 <a href="/lkml/bounce/2005/12/9/255">[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/2005/12/9/202">First message in thread</a></li><li><a href="/lkml/2005/12/9/202"> hawkes@sgi ...</a><ul><li class="origin"><a href="">Paul Jackson</a></li><li><a href="/lkml/2005/12/10/58">Ingo Molnar</a></li><li><a href="/lkml/2005/12/10/60">Ingo Molnar</a><ul><li><a href="/lkml/2005/12/12/192">"John Hawkes"</a><ul><li><a href="/lkml/2005/12/13/26">Ingo Molnar</a></li></ul></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">Fri, 9 Dec 2005 16:06:47 -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: [PATCH] -mm tree: broken "dynamic sched domains" and "migration cost"</td></tr></table></td><td></td></tr></table><pre itemprop="articleBody">> (5) Besides, the migration cost between any two CPUs is something<br />> that can be calculated once at boot time and remembered<br />> thereafter. I suspect the reason why the algorithm doesn't do<br />> this is that an exhaustive NxN calculation will be very slow for<br />> large NR_CPUS counts, which explains why the calculations are<br />> now done in the context of sched domains.<br /><br />Agreed - I too suspect that this a form of compression, both of<br />computation costs and data size. We save space and time by not<br />calculating the full N * N matrix, where N is num_onlinecpus(), but<br />just the sched domain sub-matrices.<br /><br />In theory, I would think that we should -not- compress based on sched<br />domains, because:<br /> 1) these are (now becoming) dynamic, and<br /> 2) they don't reflect the "natural" basis for such compression,<br /> which would be hardware topology based, not sched domain based.<br /><br />Rather we should compress based on the topological symmetries of the<br />hardware system. Of course, this is an ARCH specific characteristic,<br />or even platform specific.<br /><br />Perhaps we could provide an ARCH specific routine that would map any<br />ordered pair <cpu0, cpu1> of cpu numbers to a canonical pair, such that<br />the canonical pairs were "about as far apart, for that system<br />topology", but potentially much fewer in number than the entire N * N<br />space, and a smaller maximum value of the largest cpu number returned.<br />The default routine would be the identify function, which would work<br />fine for ordinary sized systems.<br /><br />A second ARCH specific routine would return the largest value M<br />canonical cpu number that would be returned by the above routine.<br />The distance array could be dynamically allocated to M**2 size.<br />The default routine would just return the highest online CPU number.<br /><br />These 'canonical cpu pairs' would replace the sched domains as the<br />basis for compression.<br /><br />Then one time at boot, for each possible pair of online cpus, map that<br />pair to its canonical pair, and if not already done, compute its<br />migration cost. For example, if on the current systems topology, cpu<br />pairs <3,5> and <67,69> are pretty much the same distances apart, the<br />"canonical" pair for both these might be <3,5>, and only that pair<br />would have to be actually computed and stored. Everytime the software<br />using this wanted results for <67,69>, it would get mapped to <3,5> for<br />resolution.<br /><br />In the extreme case of a big NUMA system with an essentially homogeneous<br />topology (all cpu-cpu distances the same), all <cpu0, cpu1> pairs where<br />cpu- != cpu1, could be mapped to same canonical pair <0, 1>.<br /><br />-- <br /> I won't rest till it's the best ...<br /> Programmer, Linux Scalability<br /> Paul Jackson <pj@sgi.com> 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: 2005-12-10 01:09 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>