Re: depend take 2

[Martin Pala]

Agree. When i though about your proposal (dependency matrix) in the 
past, i found, that best way could be to paralelize start-up process as 
much as possible => make one thread for every independent process 
branch/subtree to start/stop them. In the case that the branch has two 
or more even processes dependant on one process, it could be possible to 
start them at once. I hope following ascii-art will explain its princip:

                / p4 - p8
- p1 - p2 -
               \ p5

- p6 - p7

-p9


So the fastest way to start it will be:

1.) start three independent threads for taking care about p1, p6 and p9 
branch (all threads will continue to work indepently until thay 
start/stop all processes). The dependency layer should be completely 
independent on atomic start/stop methods (control.c stuff)
2a.) continue to start p2
2b.) continue to start p7
2c.) finished
3a.) start two threads to take care about p4 and p5 branch
3b.) finished
4a.) start p8
4b.) finished
5a.) finished

So following "algorithm" is as fast as possible and won't block whole 
process if there will be some problem. By the way, there are yet another 
two options:
I.) it could be possible to independantly monitor processes by main 
thread on the background, but there must be some mechanism to 
synchronize start-up progress with monitoring, so monit should know, 
that some processes are under "maintanance" and that it doesn't make 
sense to try start/stop them until present work will be done.
II.) just block main monit thread and wait until all the others 
(start/stop of processes) will finish

Martin
        
     
Jan-Henrik Haukeland wrote:

> Anyway, I still think that moint should suspend itself as litle as possible in
> the start program phase, max 10 seconds and if that doesn't work it should
> break out of the start-program+depend loop with a warning message in the log.
> Besides it will retry again on the next cycle anyway. This is better than to
> (worst case) sleep Run.polltime for a program. Monit should instead then go on
> to the next program that needs to be started.
>
> Jan-Henrik
>
>
>  
>
> > > Besides if a program does not manage to create a pidfile in 10 seconds
> > > something is wrong.
> > >
> > >      
> > Not true. If you are doing some kind of connection testing, it may well
> > take a full poll cycle to come up.
> >    
>
>
>
>
> _______________________________________________
> monit-dev mailing list
> address@hidden
> http://mail.nongnu.org/mailman/listinfo/monit-dev
>