Texture resolution best practices

After seeing countless comments about the importance of keeping texture resolutions low to prevent lag (on download and in video memory), I've started the habit of uploading the same texture at different resolutions so I can pick the most appropriate size. I only upload one fairly high resolution at first, just to make sure it looks right inworld, but then once it is final, I will usually do a 64x, 128x, and 256x. Depending on what it is for, I might do a 32, 512, or rarely even a 1024. In addition to a descriptive name (with keywords I'm likely to search for), I also specify the resolution in the name.

Then, when I build, I can easily find the most suitable resolution. But a thought occurred to me that the only thing worse than having to download a big texture on a small prim is having to download a small texture when you've already downloaded a bigger version. Am I right in assuming that when I make a suite of furniture (or whatever), if I've already used a 256x256 on one piece that it is probably better to go ahead and use that same one on another piece than it is to use a smaller one? This assumes, of course, that the different bits of furniture will generally be used together.

Yes.

And, another thing that I am fairly sure is right just on the basis of general understanding of computers and networking...

If you in fact need 16 256x256 textures for a room your working on, using a single 1024x1024 texture, cut into 16 panels, and appropriate offsets used to get what you want is probably significantly more efficient from a texture load standpoint than using 16 individual textures. I do this, for instance, when I'm putting up a zillion snapshots on a display board, putting up textures in my store, etc.

I've always been unsure of this method (multiple celled pictures in one massive texture) from the standpoint of the video card...

wouldn't the video card have to load ALL or the texture for each usage, even if that particular instance only used a fraction of that master? or does it (as assumed above) merely reference one loaded texture for all instances?

I know it can be nice from a loading perspective in that the entire scene or object would load simultaneously.

there's also a question of bleeding from one cell to another, making texture edges take on elements from sibiling cells, how strong is this effect with LL's jp2 implementation?

Texture size on the graphics card is determined by the number of pixels, eg.
1024x1024 = 1048756 pixels
256x256 = 65536 pixels

65536 x 16 still equals 1048756 pixels
So its the same either way.

The disadvantage with the combined image is it cant be unloaded from memory when some portions of it are out of view. If you decide to get rid of half the objects the full image still has to be loaded.

It could be possible that the jpeg compression or pallet would work better, but it could also be worse. Card memory decompresses it anyway so makes no difference there.

Saves a few cents in file uploads I guess.

More than that, it saves on network traffic. Instead of 16 separate calls to the asset server to deliver 16 separate assets, there's only one.

Also, you have the advantage that all the panels will load at the same time. If they're separate images, they'll all trickle in at random times, but if they're one, they'll all appear at once.

You're right, though, that if you don't have enough surfaces in the scene to display all the frames, then it's a waste. For a furniture set that is intended to stay together, as the OP described, it will work great. But if it's for a bunch of random stuff that wouldn't go together, you're better off breaking it down into separate textures.