1) You are correct. Auto-saving will save the data in memory, but doesn't change what's in memory in any way.
2) When auto-save saves, it takes the data in memory and saves it into the file. If that file already exists, it is overwritten. The contents of any file already on disk are not taken into account - the new file is just written over top of it (with the contents of what's in memory at that exact moment).
3) When you reload any of this data in to memory, your reloaded data will be the exact copy of what was in memory at the time of save.
4) If you load, the "Max samples to keep in memory" is disregarded until the next sample is collected. As long as you're not collecting data, this setting isn't an issue. As soon as you collect any data, however, any old data in memory will be discarded if your samples in memory exceeds the setting.
Now, let's get a bit more to the *guts* of what you're trying to accomplish.
If your save file name is set to $host $date $hour (and has a $hour factor in it), then you can set your max samples to include in memory to something as low as an hour and not lose any data. This isn't what I recommend, however.
I would recommend keeping a day's worth of data in memory. At 2.5 seconds, this is 34560 samples - or roughly 1.3 megabytes of collected data. It may be that you don't have this much free memory - in which case you might set it to something lower.
You want to keep enough data in memory so you can see a pretty good overview of what was happening - without having to load each file individually. I like to have 48 hours in memory at a time - so I'd actually set mine to 75000 or so (actually, I use 2.5 second and run mine at 200000 - but that's more than most people need). I'm not sure the constraints you're going to run in to with memory, but 1.5 megs is a drop in the bucket these days <G>.
Why do you want more data in memory? Well, you can only graph what you have in memory, so if you only have an hour worth of data in memory, then your time graphs can only show an hour's worth of data - which is useful, but isn't enough data to show trending.
Note that 1500 samples is only about 60K of data - pretty much nothing on a modern system. I'd strongly recommend having more data in memory at a time.
Feel free to follow up with any additional questions / recommendations.
- Pete