Jim Fulton wrote:

On Mar 29, 2007, at 10:11 AM, Jürgen Kartnaller wrote:

Jim Fulton wrote:
On Mar 29, 2007, at 3:51 AM, Jürgen Kartnaller wrote:

Jim Fulton wrote:
On Mar 28, 2007, at 6:37 PM, Benji York wrote:
Jim Fulton wrote:
On Mar 28, 2007, at 6:28 PM, Benji York wrote:

threading.Condition :)

Yeah, that. :)

What was the time to return the first request after the fix?

I didn't stick around for the post-Linux-kernel-recompile testing, but when we hacked the Python standard library to use sleep(0) all the time (emulating an infinite granularity timer), it went down to about 10 seconds. A pretty nice improvement. ;)
And the original time was 15 minutes! :)

My test environment on MAC OS-X is loading my frontpage in 2:50 min after restarting zope. with your fix and the one I made in a ZODB branch using nanosleep is went down to 27 sec.
I'm confused. Was the 27 seconds with both changes together? Or did each change individually lead to a 27 second time?

Each canche individual.
Both have the same result.

Cool. They both address the same problem individually, at least in theory, but I'st nice to know the 2 changes were tested separately.

Of course the nanosleep version is not elegant and brings more CPU load because I set the sleep to 100 us and therefore need to do a lookup for data every 100us.

Because nanosleep didn't worked reliable on ubuntu (sleep time varies from 100ns to 10ms) we changed the kernel to have a minimal sleep time of 1 ms. Now our application "warm up" time decreased to about 10 minutes (from about 30 or more).

I think we will use your new version next week on our production servers.


For more information about ZODB, see the ZODB Wiki:

ZODB-Dev mailing list  -  ZODB-Dev@zope.org

Reply via email to