Re: "optimal" thread number on SMP architecture

[Howard Chu <hyc@symas.com>]

Pierangelo Masarati wrote:
> It appears that a rough rule of thumb yields 16 * nCPU as the optimal 
> thread number.  I've some code that tries to detect the number of 
> CPUs, we could use it to choose the default if not set in slapd.conf.  
> I'm considering a rough algorithm like "threads = min( 16 * nCPU, 64 
> )".  Would this qulify for inclusion?
My instinct is to go with something more like 16 * sqrt( nCPU ); I have 
yet to see an SMP machine that provides performance scaling linearly 
with the number of CPUs in threaded apps. And there's also the worry of 
having all of the process' address space consumed by thread overhead.

I would rather see some mechanism for running multiple slapd processes 
from the same config and databases. That would reduce some of the 
concerns about process address space bloat (although I suppose in a 64 
bit machine that shouldn't be an issue) as the total number of threads 
increases. Ideally each process would have its own separate environment 
cache, though all would share a single lock table.

--
 -- Howard Chu
 Chief Architect, Symas Corp.  http://www.symas.com
 Director, Highland Sun        http://highlandsun.com/hyc
 OpenLDAP Core Team            http://www.openldap.org/project/