I'm using an N900 with kernel power v49.
I've downloaded titan's overclocking profiles and since the first day I started using them, I was uncomfortable with some of the settings.

Namely, the UP_THRESHOLD values and the SAMPLING_RATE values.

If you look at most of those profiles, what they did is they lowered the UP_THRESHOLD and also HALVED the SAMPLIG_RATE that you would have with stock settings.

IMHO, those should go in opposite directions, i.e. if you lower the UP_THRESHOLD you needn't to SAMPLE as often. To me this is logical.

As a result, your device will be spending 2x more processor time on the sampling task, to know if it should scale up the processor frequency or not. Besides that, the processor frequency will be scaled up before it reaches critical load, so there's significant waste of energy there.

Lately I've been experiencing some less than desired performance with I do overclock. I try to run overclocking only when I am about to start a task that makes the phone slugish, like Kasvopus or any Twitter client. So I've been thinking why this could be happening.. and decided to change these overclocking profiles!

To my honest surprise, the phone performs much much better now.
So here's to you all, community! A tip you might make good use of.


PS.: your profiles are located in ~/.kernel/myprofile1 etc...

The ondemand governor jumps up to max whenever any significant load is applied.

IIRC the threshold is used for frequency step-down.

how about battery effect?
i believe this will lower your battery life

still someone need to do wide tests to know what are the settings which will improve the speed, others will improve battery, or some where between them

The up_threshold is used for any transition (up or down).

Only when sampling_down_factor (which we don't have) is other than 1 the decision to lower the frequency can happen at a lower rate than that of raising the frequency.

Still, the fact that we have so little need for lots of frequency steps from 250 to $DESIRED_MAX_FREQ on our device plus the fact that you can have ondemand jumping from 250 to max whenever you do something (launching anything on the N900 sends CPU sky high) means that all you need is 250(media player while locked), 500(locked with processes running) and your max freq(daily tasks), with sampling rate of something like 1000, or 100 if you can wait for the jump from 500 to max.

@Hurrian,

Agree. I currently have, under KP50, only { 250 500 805 } active. It sort of feels "right" to keep the number of options small to avoid jumping around (even though the transition is really quick).

At the moment I have up_threshold 95 and sampling_rate = 300000 (us), meaning (at least in theory) that it will only jump up when really necessary, but quickly enough.

This is against what OP said, but I don't think anybody is right. In essence, sampling rate and up_threshold are independent of each other.

Whether it is stepping up or down is irrelevant, agree?
The most important point is excessive sampling, coupled with 10-25% of CPU load being unused.

@carloxic,

The so-called "sampling rate" is actually a "sampling latency", i.e. it is actually the TIME between samplings, measured in micro-seconds.

Meaning, a smaller value of sampling rate will actually mean MORE sampling.

I guess I still need to confirm that looking at the kernel, because it may be a bit confusing.

@reinob,

Yes, that's consistent with what carloxic said in the OP. Therefore, if we want the gouvernor to be more sensitive (lowering UP_TRESHOLD), it makes sense to do less sampling (*increasing* SAMPLIG_RATE) so that the CPU doesn't spend to much time in the sampling.

@Tofe,

Whatever. Maybe I'm not too awake today. But to me "more sensitive" implies lowering up_threshold (OK on that), but also "more sampling", i.e. *lowering* sampling_rate, i.e. more frequent checking, i.e. more sensitive