RE: RE: 11g RedHat 5 and Hugepages
Date: Thu, 14 Oct 2010 18:36:25 -0400
I think Tanel's guess at the configuration gets pretty close to the truth. How the application is structured at the app and web server is beyond my control, but we will not be using MTS, and jdbc connection pooling will be used. Response time will be critical, hence the massive RAC cluster and the high interest in hugepages for a 70+ gb sga.
From: Greg Rahn [mailto:greg_at_structureddata.org] Sent: Thursday, October 14, 2010 5:33 PM To: Tanel Poder
Cc: Crisler, Jon; Oracle-L_at_freelists.org Subject: Re: RE: 11g RedHat 5 and Hugepages
My point is that it isn't a per app server connection problem. I know one customer who thought that just 2 connections per app server was ok, but they had 5000 app servers. Needless to say their db system was very unstable.
From experience I've seen a 10x reduction or more in db connections result in 3x or more throughput gains with reduced transaction times. The impact of such a large number of processes is very significant on the kernel and scheduler. For whatever reason, apps people think they need way more connections that is really necessary. Sometimes it's a result of "connection squatting" (holding connections for longer than the required db requests take), so their solution was simply to add more. Similar to how some fix cursor leaks - they just set open_cursors so some crazy large number so the error messages stop.
I'm hesitant to believe that one needs a 130:1 (75000/576) db connection to db CPU ratio to achieve the desire throughput if the db connections are actually doing work. In reality a 2:1 or maybe 4:1 ratio should be more than enough for connection friendly applications.
On Wed, Oct 13, 2010 at 11:06 PM, Tanel Poder <tanel_at_tanelpoder.com> wrote:
> Well it depends of course, if you have 1000 application servers in a (HPC)
> farm all directly connecting to that database, then 75 connections per app
> server doesn't sound that much.
> MTS is a memory-saver, but not CPU saver, especially when fetching/loading
> lots of data. If you have the memory to afford all these dedicated
> connection processes, it'll be faster (and less CPU hungry) way to use
> dedicated servers...
> On Thu, Oct 14, 2010 at 6:13 AM, Greg Rahn <greg_at_structureddata.org> wrote:
>> There may be issues with MTS but no where near the issues of having an
>> application that needs 75k db connections. This must be one poorly
>> written application.
>> I can't imagine how much CPU the dispatchers would burn for that many
-- Regards, Greg Rahn http://structureddata.org -- http://www.freelists.org/webpage/oracle-lReceived on Thu Oct 14 2010 - 17:36:25 CDT