[sldev] Texture caching/bandwidth/performance how to proceed

[Robin Cornelius]

Hi everyone,

We had some more discussion at Rob's hours yesterday and the basic
problem is there are multiple issues here and we seem to be going over
the same topics in circles without really achieving very much. (In
fact Rob believes this sounds like the March 2007 discussion all over
again).

Rob proposed we create a bof[1] Working Group, and to me this sounds
like it needs some wiki pages to lay out the issues. I am happy to try
to pull some stuff together, but without sounding like a complete
moron, where should we put these pages on the wiki?

What i propose is that we firstly try to subdivide these issues and
although they are related there seems to be some funtimental top level
categories that are :-

* Performance
* Amount of downloaded data

May be more and sub issues of those too.

The original discussion was purely a performance due to decoding of
cached j2k textures. But other very valid sub points have cropped up
including issues of the amount of downloaded data. Various suggestions
have been made about storing more data, uncompressed data etc all of
which many have consequences .Other things include not using the
separate texture headers and files, storing in alternative formats  We
could also do with defining what we could do and methods of measuring
its effects in quantitative terms NOT qualitative for performance/data
download amounts etc.  But we also have to factor in to account
political views (the angry villagers with pitchforks analogy).

So what i would like to do is :-

* Define issues with the texture caching system (with the scope also
including download amounts due to textures)
* Define and discuss possible solutions, alternatives etc.
* Get some linden input on what they are doing internally, is the
texture cache being rewritten? or planning to be as part of the new
texture pipeline, is this worth wile and will it provide decent
feedback to the lindens?
* If anyone has time, play with some possible solutions and measure performance.
* if we are to measure performance, define how on earth to do this.

Regards Robin

[1] http://en.wikipedia.org/wiki/Birds_of_a_Feather_(computing)


<begin flaming>