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: Trond Myklebust: Re: NFS client latencies</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 Trond Myklebust" href="/groupie.php?aid=122" /><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/2005/3"> [Mar]</a> 聽 <a class="nb" href="/lkml/2005/3/31"> [31]</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/2005/3/31/113" onclick="this.href='/lkml/headers'+'/2005/3/31/113';">[headers]</a>聽 <a href="/lkml/bounce/2005/3/31/113">[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/2005/3/29/301">First message in thread</a></li><li><a href="/lkml/2005/3/31/23">Andrew Morton</a><ul><li><a href="/lkml/2005/3/31/29">Ingo Molnar</a><ul><li class="origin"><a href="/lkml/2005/3/31/121">Trond Myklebust</a><ul><li><a href="/lkml/2005/3/31/121">Trond Myklebust</a><ul><li><a href="/lkml/2005/3/31/139">Ingo Molnar</a></li></ul></li></ul></li></ul></li></ul></li></ul><div class="threadlist">Patch in this message</div><ul class="threadlist"><li><a href="/lkml/diff/2005/3/31/113/1">Get diff 1</a></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">Subject</td><td class="rp" itemprop="name">Re: NFS client latencies</td></tr><tr><td class="lp">From</td><td class="rp" itemprop="author">Trond Myklebust <></td></tr><tr><td class="lp">Date</td><td class="rp" itemprop="datePublished">Thu, 31 Mar 2005 06:58:24 -0500</td></tr></table></td><td></td></tr></table><pre itemprop="articleBody">to den 31.03.2005 Klokka 09:30 (+0200) skreiv Ingo Molnar:<br /><br /><br />> > And presumably that list-based code rarely hits the worst-case, else<br />> > it would have been changed by now.<br />> <br />> that was my other point in a previous mail: most write benchmarks do <br />> continuous append writes, and CPU overhead easily gets lost in<br />> network <br />> latency.<br /><br />Well. Most _good_ write benchmarks go through a variety of scenarios.<br /><br />The problem with the radix-tree is that in most cases you are swapping a<br />linear scan through only the dirty pages (most of which you want to try<br />to add to the list anyway) for a linear scan through the entire list of<br />pages.<br />Damn! I've forgotten an extra _obvious_ optimization that might help<br />here... See the attached patch.<br /><br />> Also, considering that a good portion of the NFS client's code is still <br />> running under the BKL one would assume if the BKL hurts performance it <br />> would have been changed by now? ;-)<br /><br />Funny you should mention that... Chuck happens to have spent the last<br />fortnight working on removing our BKL-addiction. ;-)<br /><br />Cheers,<br /> Trond<br /><br />-- <br />Trond Myklebust <trond.myklebust@fys.uio.no><br /> fs/nfs/write.c | 28 +++++++++++++++++-----------<br /> 1 files changed, 17 insertions(+), 11 deletions(-)<br /><br />Index: linux-2.6.12-rc1/fs/nfs/write.c<br />===================================================================<br />--- linux-2.6.12-rc1.orig/fs/nfs/write.c<br />+++ linux-2.6.12-rc1/fs/nfs/write.c<br />@@ -539,12 +539,15 @@ static int<br /> nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)<br /> {<br /> struct nfs_inode *nfsi = NFS_I(inode);<br />- int res;<br />- res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);<br />- nfsi->ndirty -= res;<br />- sub_page_state(nr_dirty,res);<br />- if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))<br />- printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");<br />+ int res = 0;<br />+<br />+ if (nfsi->ndirty != 0) {<br />+ res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);<br />+ nfsi->ndirty -= res;<br />+ sub_page_state(nr_dirty,res);<br />+ if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))<br />+ printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");<br />+ }<br /> return res;<br /> }<br /> <br />@@ -563,11 +566,14 @@ static int<br /> nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)<br /> {<br /> struct nfs_inode *nfsi = NFS_I(inode);<br />- int res;<br />- res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);<br />- nfsi->ncommit -= res;<br />- if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))<br />- printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");<br />+ int res;<br />+<br />+ if (nfsi->ncommit != 0) {<br />+ res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);<br />+ nfsi->ncommit -= res;<br />+ if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))<br />+ printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");<br />+ }<br /> return res;<br /> }<br /> #endif</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>