Linux: concerns over systemd [OT]

Jimmy Hess mysidia at gmail.com
Thu Oct 23 02:48:51 UTC 2014


On Wed, Oct 22, 2014 at 1:31 PM, Barry Shein <bzs at world.std.com> wrote:
[snip]
> The unix community has exerted great amounts of effort over the
> decades to speed up reboot, particularly after crashes but also
> planned. Perhaps you don't remember the days when an fsck was
> basically mandatory and could take 15-20 minutes on a large disk.
>
> Then we added the clean bit (disk unmounted cleanly, no need for
[snip]
> And you whisk all that away with "it's not really clear to me that
> 'reboots in seconds' is a think to be optimized"????

False dilemma.
Optimizing reboot time down from 20 minutes  to  1 minute is a
significantly meaningful improvement;  it's literally a 85%  reduction
in time spent during each boot process  from the original time.

Reducing boot time from 20 minutes to  10 seconds is not significantly
better than reducing it to 1 minute.


A different choice of tradeoffs is more appropriate to different kinds
of systems, depending on their use case  (Desktop vs Server)!

Especially, when the method of reduction  is subject to diminishing
returns and increasing fragility or increasing complexity -- greater
risk that something is breaking or more potential for unreliability is
introduced into the startup process.

Also, you may very well spend more time booting your system in order
to troubleshoot,  the fact that some applications are starting up in
an unexpected order  resulting in some issue.

> To me that's like saying it's not important to try to design so one
> can recover from a network outage in seconds.

If you need to ensure that a service is not disrupted for more than
seconds,  then reboot is not the answer.     It is some form of
clustering.

Reboot as a troubleshooting procedure is for desktops.
10 seconds from power on to user interface for desktops, will
meaningfully improve the user experience,  but not for servers.


For servers, you ideally want to take the misbehaving node out of
service and let its failover partner takeover.

-- 
-JH



More information about the NANOG mailing list