I'm about to work on an improvement to my resource managers by adding on demand resource streaming.

I read these articles on texture streaming for some ideas:

http://udn.epicgames.com/Three/TextureStreaming.html
http://home.comcast.net/~tom_forsyth/blog.wiki.html (Visibility and priority tests for streaming assets)

I plan on starting slow and just making resources load in as needed, and when doing so, kick out any resources that haven't been used in 90 seconds.

My plan for later is to keep things in memory for a while and only kick them out if the memory usage is high enough. Only problem is knowing what that memory usage level is. For textures, would it basically be something around the Video memory? How bout meshes, or audio, and so on... That's what I'm really confused about right now.

I'd be really interested in responses to this too. Specifically (and hopefully adding to the info the OP's request), if you have an object that is a fair distance away and needs to stream in a texture, do you reserve a texture of the required number of MIP levels and resolution of the fully loaded texture and then continue filling up MIP levels as you get closer/require more detail?

If so, then if you don't ever need more than 1 or 2 MIP levels, aren't you wasting texture space? (there will be 4 or 5 MIP levels available and waiting in that texture with no use).

The only other way I can see it working is it you have a pool of different-sized textures with 1 MIP level, a pool with 2 MIP levels, a pool with 3, and so on and as you get closer to textures, you switch all the already-loaded MIP levels over to the next available '+1 MIP level' texture and load the next MIP level. But then this might introduce some latency with swapping texture data around...

Any thoughts on the OP's question and/or this?

I remember one discussion where someone said that it's slower to load only up to some number of mips, and as you get slower, to load more mips because you are basically creating a whole new resource.

It might be a bit of a waste of memory, but you allocate all mips in the texture you are streaming, but actually only load up to the one you need at that moment. The benefit comes from gradually loading in only what you need instead of all at once.

There's even more added complexity when you start adding prediction and priorities to resources. Like, how do you assign the different priorities, and at what memory usage do you start kicking out the lower priority resources. You could say, oh just use video memory, and track how much memory the textures are using, but then you also have your meshes, shaders, and a bunch of other things taking up video memory. Now you have to balance those objects too, do I allocate some percentage of memory for textures, some percentage for meshes? I don't even know how to begin figuring out memory usage for other things like shaders or how any of that stuff works.

Then there's sound and other resources. You have a computer with some varying amount of memory and some varying processes running taking up random amounts of memory. On consoles it's all set in stone more or less, but on PC, how would I determine if I'm using too much memory for sound effects and so forth. And video memory isn't fixed on PC's either since different cards have different memory.

For a much simpler game a simpler resource manager would probably do just fine...