[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: REPLACE_BROKEN_YIELD [was Re: Comparing OpenLDAP on Linux 2.6/Sol 10 ]

--On Tuesday, January 17, 2006 3:37 PM -0800 Ron Grant <rgrant@shuttle.net> wrote:

What does this value do, and is it best to leave it set on x86 Linux
v2.6.xx kernel?

Other than upgrade to 2.3.18, which fixes the mess...

It tried to fix the issue with sched_yield() that the linux 2.6 developers introduced, but the result was worse than how it behaved without this fix. The configure test also didn't differentiate between 2.4 and 2.6, so the problem would manifest with OpenLDAP built on either kernel (2.3.14-2.3.17).

This value appears to be set by "configure". I'm trying to build an RPM
so would have to modify that value after "configure", but I'd like to
know what it's going to affect first. Stability is more important than
speed, but better speed would be nice. Should I just setup a test and see
if it works on my server?

I ran 1.5 million unique searches on my test server, a 2.4.Ghz P4 (with a
maxtor IDE HD, I might add) from fifteen different clients and finished
in 24 minutes. This is 1000 completed searches per second, more than
enough for my needs, but hey...... "fast = good"...no?

As a comparison, with 12 clients querying, I maxed out my 100MB ethernet with the size of entries I'm using at 8,664 searches/second, and by dropping the entire entry and using just a single attribute returned, got around 14,400 searches/second. That's on a dual CPU 3.0ish GHz P4. Of course, rates depend on many different things.


Quanah Gibson-Mount
Principal Software Developer
ITSS/Shared Services
Stanford University
GnuPG Public Key: http://www.stanford.edu/~quanah/pgp.html