Hi Scott: Something else for you to play with:
http://www.mail-archive.com/[email protected]/msg03614.html Cheers, Bernard On Thu, Feb 25, 2010 at 3:12 PM, Scott Dworkis <[email protected]> wrote: > finally had a chance to finish up an rrdcached config yesterday. > > comparison: > > attached some 1 week graphs showing resource usage of proc/sys/vm tuning > before and rrdcached after last night. i reverted my proc/sys/vm > customizations to the defaults and set up rrdcached to what i felt was > comparable to the proc/sys/vm settings i used, basically a 2 hour write-back > schedule (i.e. -w 7200 for rrdcached as opposed to > proc/sys/vm/dirty_expire_centisecs to 720000). > > the average cpu and load is just a hair higher, the 5 minute max cpu and > load are about 50% higher (i.e. rrdcached is a little choppier). disk load > seems a little lower and wider... probably equivalent on average. memory > that was used by fs cache is now used by rrdcached of course... i think this > means the data exists in 2 places, fs cache and rrdcached, which doesn't > seem strictly optimal but is probably benign since the os would reclaim fs > cache if it ever needed it for other processes... though perhaps there is > some overhead there (probably trivial). > > you might notice the triangle memory leak in gmond... i have a thread going > elsewhere for that and will have a look with valgrind. > > > > someone may want to update the guide: > > http://sourceforge.net/apps/trac/ganglia/wiki/rrdcached_integration > > i found that following that guide literally didn't work... i needed to > specify -l explicity. even though the socket file is created and writable > by all, i was getting permissions errors without -l. > > also running make on rrdtool-1.4.2 surprised me a little when it called > sync, which took a long time on my heavily cached ganglia server. i shut > down gmetad to try to get make to finish a little faster. > > > > in the end the tradeoff with rrdcached is that you isolate the caching > policy to just ganglia and not the wider system, which seems like a > good/safer thing. like Kostas noted, besides customizing init.d scripts and > default ganglia php files to make them rrdcached aware, you also have to > configure rrdcached for anything else that intends to access up to date data > in the rrds... a little more config work. > > > -scott > > (i replied to the digest here... hopefully i didn't screw up threading > somehow) > > >> ------------------------------ >> >> Message: 5 >> Date: Fri, 19 Feb 2010 11:08:41 -0800 >> From: Bernard Li <[email protected]> >> Subject: Re: [Ganglia-general] experiences scaling to 150k metrics >> To: Scott Dworkis <[email protected]> >> Cc: [email protected] >> Message-ID: >> <[email protected]> >> Content-Type: text/plain; charset=ISO-8859-1 >> >> Hi Scott: >> >> On Thu, Feb 18, 2010 at 7:41 PM, Scott Dworkis <[email protected]> wrote: >> >>> i tried the tmpfs thing (actually /dev/ramN), but decided syncing back >>> and >>> forth to disks to preserve data across reboots was suboptimal. ?never >>> tried >>> rrdcached. >> >> Yeah that's one of the downside of the tmpfs route. Definitely give >> rrdcached a try and report back your experiences. >> >> Cheers, >> >> Bernard >> >> >> >> ------------------------------ ------------------------------------------------------------------------------ Download Intel® Parallel Studio Eval Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta. http://p.sf.net/sfu/intel-sw-dev _______________________________________________ Ganglia-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/ganglia-general
