| 1 |
Rich Freeman <rich0@g.o> writes: |
| 2 |
|
| 3 |
> On Fri, Aug 17, 2012 at 3:29 AM, Luca Barbato <lu_zero@g.o> wrote: |
| 4 |
>> It can be not so tiny, surely busybox+openrc gives a better gain in many |
| 5 |
>> cases. |
| 6 |
>> |
| 7 |
> |
| 8 |
> I suspect that it will depend greatly on what services you're running, |
| 9 |
> and what order they happen to start in, and what you care about. In |
| 10 |
> theory slamming the kernel with a ton of processes will allow it to |
| 11 |
> manage its queues better with a fuller understanding of demand. |
| 12 |
> systemd can potentially short-cut this a bit further since it can |
| 13 |
> consider a dependency resolved if nothing more than a socket is |
| 14 |
> created, which is a clever trick (I have no idea how well it works out |
| 15 |
> in practice, though I have used a .socket service once and that worked |
| 16 |
> out fine (with the caveat that the first connection fails)). |
| 17 |
|
| 18 |
Yeah, this is a brilliant idea. |
| 19 |
|
| 20 |
A side question, what is the percentage of the dependencies that is |
| 21 |
merely a socket? |
| 22 |
|
| 23 |
> Where I saw the bigger performance difference between openrc and |
| 24 |
> systemd was in shutdown. Systemd shuts down REALLY fast, and I've |
| 25 |
> noticed it tends to actually kill ssh sessions rather than leaving |
| 26 |
> them unresponsive (plus ssh dies even faster so that makes it seem |
| 27 |
> subjectively faster - but the difference in full shutdown is still |
| 28 |
> real). Obviously people don't care as much about shutdown |
| 29 |
> performance, but on a laptop (where supposedly event-driven inits |
| 30 |
> should shine) I know that waiting for a full shutdown before packing |
| 31 |
> up can be painful at least in the land of Windows. |
| 32 |
|
| 33 |
Yeah, that is a problem. We've got to learn from systemd here. Not sure |
| 34 |
of the mechanism yet. |
| 35 |
|
| 36 |
> Just food for thought - it seems odd that parallel startup should have |
| 37 |
> any issues at all, provided that service scripts don't terminate until |
| 38 |
> the service is functional and all dependencies are specified. If |
| 39 |
> either of those aren't true then you'll have race conditions. |
| 40 |
|
| 41 |
OpenRC doesn't have protection against weak (in the sense of use / |
| 42 |
order) circular dependencies. I guess that is the source of this |
| 43 |
problem. |
| 44 |
|
| 45 |
Cheers, |
| 46 |
Benda |
Archives