[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Request for feedback and/or participation on https://fedoraproject.org/wiki/Features/PowerManagement



Matthew Garrett wrote:
> On Wed, Feb 04, 2009 at 05:04:56PM +0100, Phil Knirsch wrote:
> 
>> Network card speeds and harddisk spindowns is a bit tricker imo, though 
>> we probably could enable a higher powersave mode for harddisk by default 
>> at least in case a user wants to save power. Some of this is/will be 
>> addressed by the tuned i'm working on which monitors the usage of 
>> network and harddisk devices and dynamically adapts the settings to the 
>> current use.
> 
> The problem with network card retraining is that you lose link for a 
> significant period of time, which is a pain (especially if you want to 
> retrain upwards because you're trying to send a pile of data...). Hard 
> drive APM has some irritating behavioural aspects - a lot of drives will 
> unload the heads very aggressively, which can cause you to hit SMART 
> thresholds in 6 months or so of normal use. We probably want to talk to 
> vendors about that.

I've tried, but with no luck.  Maybe I can get Ric to help.  :)  FWIW my
apple hardware running an apple OS also far exceeds the thresholds.

But yes, any sleep timeout should probably be set to something
appropriate that won't overrun the rated spec over 5 years or something...

> Hard drive spindown is an interesting problem that I've been looking 
> into. There's a lot of published work on adaptive algorithms for this, 
> but the most interesting is the Helmbold one. Unfortunately, it doesn't 
> actually seem to work - the behaviour is dependent on the number of 
> experts in the system, when really it should be invarient of that.
> 
>> Laptop mode and a higher value for dirty_writeback_centisecs, there i'm 
>> personally torn a bit myself. I personally do use it on all my machines 
>> (laptop and desktop), but i can understand people who are a bit paranoid 
>> about their data integrity not wanting to have that by default. But i'd 
>> personally be all for going that way.
> 
> Anyone interested in data integrity should be calling fsync() at 
> appropriate times. I'd be entirely in favour of bumping that up to 5 
> minutes or so.

Except that well, nobody calls fsync().  :)

In fact ext3 may have scared most people away from fsync!  (firefox
"bug" remember...)

OTOH ext4 will also likely expose folks playing fast & loose with their
data, because the old data=ordered + 5s journal commit time in ext3 made
the window very small.  Now w/ delalloc in ext4, things may get
interesting in this regard.

-Eric


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]