Statistics window

[Omniflux]

I have changed the statistics window for Samples / Sec etc. to use an exponential moving average instead of a rolling mean.

This increases responsiveness to large changes in the value being measured due to, for example, a change in the number of threads or the system giving more cycles to another process. The rolling mean method could take several minutes to stabilize after such a change.

The downside is this increases the jitteryness of this value again, but hopefully not too much.

[zsolnai]

Bright idea, thank you.
I don't know about the exact window size as of now, but as the characteristic of this weighting function is exponential, you may want to consider using a slightly smaller one as the strength of the older weights gets negligible in a few steps.

[jeanphi]

Hi,

Instead of a fixed filter value you should probably use min(1, elapsedtime/filtertimewindow) otherwise depending on the frequency of the calls to the update method you'll modify the filter response (and don't forget to treat the elapsedtime==0 cas like before otherwise you might blow up the computation). The added benefit will be a more commonly understandable parameter

Jeanphi

[Abel]

Thanks, I'll give it a try soon. I'm happy about the change because one of the things I use the statistics for is to see if the system is working at optimal speed and if no swapping is occurring. If there is swapping, I tend to close other programs and then watch if LuxRender's S/s value goes up; lately this didn't work too well as like you say it tended to take a long time to get closer to current speed. I'm sure there are other means to check if there is enough memory available to LuxRender but I like the direct feedback of LuxRender's statistics.

[Pilchard123]

Thanks, I'll give it a try soon. I'm happy about the change because one of the things I use the statistics for is to see if the system is working at optimal speed and if no swapping is occurring. If there is swapping, I tend to close other programs and then watch if LuxRender's S/s value goes up; lately this didn't work too well as like you say it tended to take a long time to get closer to current speed. I'm sure there are other means to check if there is enough memory available to LuxRender but I like the direct feedback of LuxRender's statistics.

What OS are you using? If you're on Windows, try task manager's (CTRL-SHIFT-ESC) processes tab, and possibly the performance tab.

[Omniflux]

Instead of a fixed filter value you should probably use min(1, elapsedtime/filtertimewindow) otherwise depending on the frequency of the calls to the update method you'll modify the filter response (and don't forget to treat the elapsedtime==0 cas like before otherwise you might blow up the computation). The added benefit will be a more commonly understandable parameter :)

I don't understand this. Do you mean I should be using min(1, elapsedtime/filtertimewindow) in place of the fixed value I am using for the coefficient α?

[jeanphi]

I don't understand this. Do you mean I should be using min(1, elapsedtime/filtertimewindow) in place of the fixed value I am using for the coefficient α?

Exactly.

Jeanphi