On 04/23/2014 05:31 AM, Lennart Poettering wrote:
On Tue, 22.04.14 10:44, Thomas Woerner (twoerner@xxxxxxxxxx) wrote:
firewalld is not the "number 1 slowest component on Fedora, right now.", but
it is plymouth-quit-wait.
No it just waits for other services to finish (as you have seen it
went down without firewalld).
Yes, but all others increased. Therefore the question: Why are other
services taking longer to start if firewalld is not started and not
installed anymore? Without firewalld the other services in the
system should start in the same time as before with firewalld
installed and started. Otherwise the calculation is just some number
and only partly related to the started service itself.
Lennart, I think you should be able to explain this discrepancy.
systemd-analyze will tell you the raw numbers how long a service needs
to start. It can provide you with an indication what is going on, but
you have to read it with a grain of salt, since it will always include
times a service just waits for another service and doesn't actually
consume CPU nor IO. Moreover, the buffer cache is a major source of
noise here, since earlier services pay a greater penalty for IO accesses
than later services. The readahead logic adds even more noise.
Ultimately this means: it's a system where performance behaviour of
services influence each other even if they don't directly
communicate. To make the data more reliable, you'd could drop the
read-ahead caches, disable excatly one service of the boot, then boot 2
times and measure the resulting total boot speed over a number of
subsequent boots. Then, reenable that one service, and disable another
one, repeat... This will tell you how much every service actually
contributes to the boot time, while staying close to the full system
where all services are enabled.
The data systemd-analyze is hence merely a trend. It indicates that
firewalld is the worst offender, and given the margin I am pretty sure
this will also be the case if you do the more accurate testing suggested
above.
Only because it adds the start times for polkit, dbus and a lot of
system things to the start time of firewalld. But not to the other
services using these. It adds it to firewalld, because it is started
very early. So either add the start times of the requires to all
services using them, or just forget about the time information in
systemd-analyze at all.
What you are naming "trend" is not even this. See above.
If I am adding "After=" lines for everything that is started
additionally, the start time of firewalld is really small.
Lennart
Thomas
--
desktop mailing list
desktop@xxxxxxxxxxxxxxxxxxxxxxx
https://admin.fedoraproject.org/mailman/listinfo/desktop
