Hi Grendel, thanks for testing it!
Hmm - from V1 15 minutes to V2 23 seconds is a factor of nearly 40! That's certainly a record so far.
I tend to get something around 20 to 25x on my quad core over here, so your measurement is fairly consistent with around double that, but your measurement with the new engine limited to a single core seems to show that there may be some further tuning possible. I'm not very sure that I will be able to do that right now though, further tuning beyond this may be fairly hard work. One thing that is very good though is that it does not choke with 8 cores running, it is actually possible with multi-threaded code to get cases where increasing the number of cores actually makes things worse instead of speeding things up, if there is a problem of too much "contention" between the cores.
Re: running out of memory - yeah since now you can more easily spew out a huge number of polygons there is also an improvement to handle out of memory conditions better, instead of just crashing or something it will show that warning when it ran out of memory when trying to complete the current task.
I guess that probably happened when you got most or all of the 8 cores happening on crunching away on those long helix surfaces, those will kind of spike to a higher level of memory consumption during their subdivision process, so getting a bunch of those all going at once will put a pretty big spike in memory consumption. It will bump back down after each one is finished and the final mesh is extracted but if you bumped up beyond your full system memory during the spike, that will be a problem.
That's kind of an interesting problem with a higher number of cores that I hadn't really considered too much.
I think you mentioned that you are running XP 32 bit? Are you already running with the /3GB switch in boot.ini? ( http://www.microsoft.com/whdc/system/platform/server/PAE/PAEmem.mspx)
If not then this may be something you'll need to use to increase the available memory for MoI to allocate when you have some longer surfaces in your model plus a high number of cores.
Anyway, it is pretty cool to go from something that took 15 minutes to less than 30 seconds on the same machine! :)
- Michael
|