tag:blogger.com,1999:blog-20038672.post251806318400197937..comments2024-03-28T00:58:29.187-04:00Comments on Robert Haas: Scalability, in Graphical Form, AnalyzedRobert Haashttp://www.blogger.com/profile/08393677427643988650noreply@blogger.comBlogger10125tag:blogger.com,1999:blog-20038672.post-6126376708605902022012-12-19T13:29:06.021-05:002012-12-19T13:29:06.021-05:00Dear Mr. Robert, please look at the papers "A...Dear Mr. Robert, please look at the papers "A more robust regression approach to estimate the parameters of super serial scalability law for noisy data" and "Mythbuster for Guerrillas" presented at CMG 2012.<br /><br />-Jayanta ChoudhuryAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-20038672.post-66647651523044692042012-04-26T02:26:05.480-04:002012-04-26T02:26:05.480-04:00I know this could be a dumb question. But I heard ...I know this could be a dumb question. But I heard somewhere that MySQL does not scale well with cores. If I need to put up a sever with multiple databases(different web sites running django, LAMP, Rails etc opting for a dedicated database server), and if I have a 24 core (2x 12 core Opteron) for database server, will Postgres scale betetr, and if so by what magnitudeRamdas Shttps://www.blogger.com/profile/13797830818307817284noreply@blogger.comtag:blogger.com,1999:blog-20038672.post-56560498210457755262012-04-11T11:55:45.343-04:002012-04-11T11:55:45.343-04:00Robert, Baron, et al.,
Come and take a gander at...Robert, Baron, et al., <br /><br />Come and take a gander at <a href="http://perfdynamics.blogspot.com/2012/04/postgresql-scalability-analysis.html" rel="nofollow">PostgreSQL Scalability Analysis Deconstructed</a>. <br /><br />Comments welcomed.<br /><br />--njgNeil Guntherhttps://www.blogger.com/profile/11441377418482735926noreply@blogger.comtag:blogger.com,1999:blog-20038672.post-54706723697265459002012-01-16T04:00:50.578-05:002012-01-16T04:00:50.578-05:00I've seen the superlinear scaling as well.
I ...I've seen the superlinear scaling as well.<br /><br />I attribute it to CPU affinity. When you have 1 pgbench thread and 1 backend process per CPU, the kernel migrates them so each thread is on the same CPU as the backend it drives.<br /><br />But when threads+backend < #CPU, the kernel tries to give each thing its own CPU, breaking up the driver/driven pairing.<br /><br />Cheers,<br /><br />Jeff JanesJeff Janeshttps://www.blogger.com/profile/09158179752309939608noreply@blogger.comtag:blogger.com,1999:blog-20038672.post-70350939926655575532011-10-23T20:35:48.823-04:002011-10-23T20:35:48.823-04:00Maybe I'm stating the obvious, but I've se...Maybe I'm stating the obvious, but I've seen similar effects due to power management features of recent hardware and kernels. I'm not sure about your configuration, so this may not apply.<br /><br />The biggest suspect is CPU frequency scaling; whenever the kernel detects that a CPU isn't active enough, it downclocks it, but there's always some lag before the clock is raised again. When you reach a stage where all CPUs in your system are mostly busy, the kernel stops downclocking and everything suddenly goes faster. Fortunately it's very easy to turn this off.<br /><br />Besides that, there are more subtle power management techniques -- PCIe controllers and RAM modules include PM features too these days.intgrhttps://www.blogger.com/profile/02974314893080470877noreply@blogger.comtag:blogger.com,1999:blog-20038672.post-77379436409579484702011-10-06T22:16:03.109-04:002011-10-06T22:16:03.109-04:00The disadvantage of the tool I created that uses g...The disadvantage of the tool I created that uses gnuplot is... it uses gnuplot, which can't place constraints on parameters. It makes no sense for there to be negative seriality or coherence, but I can't tell gnuplot not to go negative. The solution is to use R instead.<br /><br />It is possible for systems to have better than linear scalability if there is an effect of "economies of scale," that is a resource that is more efficient when shared than when used singly. I think this is relatively rare. The USL does not model this. I think this is a shortcoming of the USL model (all models are wrong, some models are useful).<br /><br />You can find another example of this here: http://mikaelronstrom.blogspot.com/2011/05/better-than-linear-scaling-is-possible.html<br /><br />I think the USL needs another parameter to reflect what I am calling "economies of scale."Baronhttps://www.blogger.com/profile/01621441847303652718noreply@blogger.comtag:blogger.com,1999:blog-20038672.post-62173424532006725952011-10-02T01:22:42.754-04:002011-10-02T01:22:42.754-04:00We will try it with OpenERP as soon it's relea...We will try it with OpenERP as soon it's released. I'm curious to see how much faster it is with OpenERP 6.1Frontware Internationalhttp://www.frontware.comnoreply@blogger.comtag:blogger.com,1999:blog-20038672.post-78719607667944541982011-10-01T19:58:01.906-04:002011-10-01T19:58:01.906-04:00This is without a patched lseek, right? If so you ...This is without a patched lseek, right? If so you should get a bit more from that once the patch lands in Linux.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-20038672.post-1984743173821939422011-10-01T04:50:58.850-04:002011-10-01T04:50:58.850-04:00It is absolutely amazing!
Can't wait 9.2 final...It is absolutely amazing!<br />Can't wait 9.2 final to see it in real action. Thnx for your work.WildRAiDhttps://www.blogger.com/profile/08433269429884087726noreply@blogger.comtag:blogger.com,1999:blog-20038672.post-59042336849636909832011-09-30T16:02:19.068-04:002011-09-30T16:02:19.068-04:00Assuming all these processes are acessing mostly t...Assuming all these processes are acessing mostly the same memory for code as well as data, and also assuming they use more memory than the L1 cache can hold they might profit from mutual "cache prefetching".Rupert Kittinger-Sereinighttps://www.blogger.com/profile/07233596493741993301noreply@blogger.com