Matt Ingenthron wrote: > Hi Sriram, > > There was discussion of this at one point, quite a while back. We > didn't do anything though > > Sriram Natarajan wrote: > >> Hi >> I have couple of questions on my apache default configuration that we >> integrate within opensolaris and it was brought to my attention that I >> raise this issue with our in house apache httpd experts. Currently, we >> provide the following tuned settings within our httpd.conf >> >> <IfModule prefork.c> >> ListenBacklog 8192 >> ServerLimit 2048 >> MaxClients 2048 >> </IfModule> >> >> Now, this is a MPM specific tuning. I am wondering, if we should move >> such mpm specific tuning within /etc/apache2/2.2/conf.d/httpd-mpm.conf ? >> >> I guess, this value was provided as a default value based on the metrics >> observed within our cool stack version of apache ? Now, within cool >> stack, there is only prefork mpm. >> > > Yes, and I think these remain from when Cool Stack was CMT systems > only. The tuning values for other architectures are probably different > (though ListenBacklog should stay the same). > > Speaking of which, where do you set the backlog on worker? If it's > defaulting to the OpenSolaris default, I don't know if that's good > (unless it's changed in S10). > > We need to have <IfModule worker.c> to set any worker MPM specifc settings. Currently, this is not the case though. >> however, in our case - web stack >> version - we do integrate both worker as well as prefork mpm. So, it >> probably makes that we provide tuned values for both these mpm. >> >> Also, should we not probably consider having different default values >> that suites best for niagara specific processors compared to intel / amd >> . Now, with respect to coming up with tuned default values - it is a big >> question on its own and I will leave it to apache project members and >> experts to sort it out and come up with the numbers :-) >> >> > > One suggestion here: we should perhaps have a few "sample" configuration > files for different systems. i.e. a 4 core X64 system, a 64-strand CMT > system, a 32-strand CMT system and a developer configuration. The end > user could then copy these into place. > > A good starting point would be some guesses, then some testing to really > determine the right settings. > Af course, you do mean that our apache project members need to make this happen based on different workloads and come up with some numbers that will be a good starting point for our customers. In any case, I wish, our IPS pkg action or SVR4 post install would allow us to recognize the hardware and place the tunings appropriately. This would seem more integrated and in that case customer does not have manually set up every node in a large data center.
- Sriram > ________________________________________ > > > webstack-discuss mailing list > webstack-discuss at opensolaris.org > http://mail.opensolaris.org/mailman/listinfo/webstack-discuss >
