On 7/12/06, Jeffrey Law <law@xxxxxxxxxx> wrote:
On Wed, 2006-07-12 at 11:51 -0600, Stephen John Smoogen wrote: > The main issue is that people want their stations to power up pretty > quickly and have a complete running system in about the same time as a > disked station. Yea. I haven't timed it, but the boot times seem comparable, at least on a small network of machines. The interesting case is how well does the system handle all the machines in the building rebooting after a power bump. This problem was largely avoided in the past simply because the servers took *far* longer to fsck their disks than they do today, fsck time dominated the time to reboot the building. So the fact that we were deep on the backoff timers for tftpboot really was a drop in the bucket.
Hmmm yes.. the diskless section I dealt with had close to 100 diskless boxes per server using a 'build each new image when you upgrade'. It took about 5 hours and lots of boxes having to be restarted to rebuiild all of them on gigabit.
These days the server reboots fast due to log based filesystems, so I'm sure the dominating factor is going to be new. Even something like the puppet bits which take a couple seconds might be significant and relevant.
THis may be an unrelated problem than what you are talking about.. in that case just let me ramble on. A fix we thought about was building a very upstream server that would make a lot of file coverage/building on the server. Box PXE boots first time gets just enough stuff to say "I am new please build me". This would put the box into a cfengine class that had the server do all the hardlinks with the master files that wouldn't change but were not 'sharable'. The box would just poll the server until this was done.. and then schedule a reboot which would put it into the end class it needed (diskless engineer, diskless admin, etc). -- Stephen J Smoogen. CSIRT/Linux System Administrator