[EMAIL PROTECTED] wrote: >> Indeed. I know of one particular installation that was installing >> 200-300 nodes at once and kickstart >> was still rather zippy (15 minutes or so), without a mirror. >> Unless you >> are seeing time problems this way, I'd go >> the kickstart route, and then if you /are/ seeing time problems then, >> come back, and there are a few ways to mitigate >> that. >> >> First off, install server mirror setup is especially easy in new >> cobbler. As an alternative >> Cobbler also has the "--server-override" parameter so all systems can >> PXE off of one box and then >> serve their content from multiple mirrors. This way you can, if you >> want, explicitly tell certain systems >> to boot off a different http tree. >> >> # establish a mirror on cobblerslave1 >> cobbler replicate --master=mastercobbler.example.org >> --full-data-sync # >> first time only, later use other flags >> # establish a mirror on cobblerslave2 >> cobbler replicate --master=mastercobbler.example.org >> --full-data-sync # >> first time only, later use other flags >> >> # set up a certain cobbler system to use mirror2, regardless of who >> actually serves it via PXE >> cobbler system edit --name=foo --server-override=cobblerslave2 >> >> I have also seen simpler things done with just a squid cache >> or mirror >> of the install tree and some clever things in %pre >> to decide which "url" line to use in the kickstart. >> >> I am fairly convinced that the problems caused by dealing with images >> (not knowing what's in the image, needing to build it, >> how do you rebuild it or build it when distributions come >> around, making >> it work with differences in hardware that the installer >> would normally account for) are not worth dealing with in most >> applications. >> >> With kickstart, you get the same advantages that a good configuration >> management system buys you -- you have a record of how >> everything got the way it was and you can easily rebuild them >> and make >> changes. >> >> With images also still transferring the image over the >> network, and the >> time spent to evaluate and process the kickstart is probably not that >> intensive. >> >> Also with images it's easy to get in the habit of not updating them >> properly because they are "golden" which is a euphemism that really >> should not be used. >> >> Especially in Fedora 7 and later (and should be in EL 6 and >> later), you >> can attach yum repos in kickstart and make sure you get all of your >> updates to all of your packages >> at install time -- no more calling yum in %post. >> >> --Michael >> > > > My use case is 8000-10000 nodes per datacenter and I need to get them up > in less than 30 minutes. These nodes use gigabit interfaces. I need to > have a centralized point to get state information for the management > software so I like the idea of using a single PXE server for boot/dhcp > and mirror the packages on other systems/subnets for load balancing. For > now I'll use kickstart and see if there are any problems. Thanks. > > > _______________________________________________ > cobbler mailing list > [email protected] > https://fedorahosted.org/mailman/listinfo/cobbler > I can't see 10,000 nodes being reinstalled all at once on a regular basis, surely you don't mean all of them at once?
A single PXE server is probably not sufficient, nor is a single dhcp. Thankfully there are ways of dealing with this (multiple next servers, cobbler can help template this with --dhcp-tag (see Wiki) and multiple dhcp servers are up to you. Staged bring up is probably a more realistic answer. _______________________________________________ cobbler mailing list [email protected] https://fedorahosted.org/mailman/listinfo/cobbler
