On Wed, 2009-01-28 at 15:26 +0200, Tomas Winkler wrote: > > I **completely** disagree on this. The non-unit index is a **horrible** > > thing because it says **nothing** of interest to the user. The only > > approach the user can then take is try which value will give him the > > desired behaviour, and it will differ across all hardware combinations. > > Therefore, it's completely _useless_. > > What I suggest is to have rich configuration for setting different > parameters so this is more flexible with different HW but in run time > only play with single number to create uniform interface for user > space. Which is completely pointless, as I'm saying, because the user has no idea what to base the decision on. This might be nice for a developer to play with this, but for a user it's completely useless because they cannot base the decision on any information. Basically, you're saying: Let us vary the blackbox parameters for you, and we'll give you a 0-5 scale on which to select. Except that the user has no idea what 0-5 mean in terms of what he's trying to do. That just means we'll end up with howtos like this: "If you have Intel hardware, and want to use VOIP, then you need to do iwconfig wlan0 power index 2, because otherwise the latency is unacceptable (if you have bad hearing 3 might be acceptable); if you're just running ICQ then 5 is fine." Can you see what I'm trying to say? > I'm not sure that you want to user jungle with all the numbers, we run > empirical tests for very long time how do you want to user space take > the decision. No, we definitely do NOT want userspace to juggle any of the numbers. But that's completely orthogonal. We want userspace to tell us what it needs, and adjust all the possible parameters based on that. We definitely do not want userspace to give us an arbitrary number it obtained by rolling a dice. > > What we really need to do is tie into the pm_qos framework and make > > latency guarantees. > > Even pm_qos doesn't talk directly in language of all PM paramters you > need to configure so this is just equivalent to my proposition and can > be staged after the legacy PM is implemented. Also PM doesn't map only > to traffic requirements. No, you're thinking too technical if you say this is equivalent to your proposition. Yes, it is similar in that userspace gives us one or two values and we select a number of parameters based on these values. But it is _completely_ different in that userspace actually has a good idea what the parameters it can tell us _do_. Unlike an arbitrary power index it can only select by rolling a dice and hoping it'll work. johannes
Attachment:
signature.asc
Description: This is a digitally signed message part
