[numerobis]

It looks like mxi merging is still single threaded... if multicore usage would be possible i think this would be a huge improvement for bigger files.

The same applies to simulens and image saving.

[polynurb]

i think for merging hdd speed is the prime factor.
the thing is with "manual" merging one can choose two hdd drives one for reading the mxi and a second for writing.
i observed a massive speed up doing it that way compared to auto-merging via network.
as i remember even 2-3GB mxi don't take longer than a minute doing it the manual way.

i hope this "target drive" option can be made available some times..

anyway i totally agree with image saving/simulense.

also i don't quite get why, when applying simulense to an image, the calculation needs to be repeated when saving out a .hdr/.exr.

[tom]

i don't quite get why, when applying simulense to an image, the calculation needs to be repeated when saving out a .hdr/.exr.

Otherwise, we had to overwrite the master buffer and you wouldn't be able to change Simulens during/after the render.

[polynurb]

Hi Tom

What i mean is eg.
Opening a mix &applying simulense
Then after the effect is applied saving 8bit is almost instant but saving 32bit takes way longer when using simulense compared to just saving 32bit without simulense.

So it seems 8bit is 'buffered' but 32bit needs to be recalculated on the fly when saving.
Does that make sense or am I missing something?

[tom]

So it seems 8bit is 'buffered' but 32bit needs to be recalculated on the fly when saving.
Does that make sense or am I missing something?

Exactly. The display itself is 8 bit buffer but you need to recalculate for 32 bit or you need to waste extra hdr buffer and so memory.

[polynurb]

Ok now I get it. It makes sense to save ram.

It is just very time intensive to deal with 3k+ images and simulense.
So I'd preference to go through the process only once even if it would cost extra ram, or let's say having the option.

Alternatively couldn't an hdr&exr be written silently to Maxwell's ram, which are only copied if the user decides later to save 32bit?

[numerobis]

i didn't think it's the network or hdd speed... from the point all MXIs are gathered in one folder on the manager and the merging starts (8x 3,5GB MXIs in this case) it took ~15 minutes til the merging was done and the images (render, object, material) were written.
Drive is a HDD RAID0 (~200-250MB/s theoretically). Manually copying the files on this drive is much faster. For the whole time of the procedure only one thread was used.
But maybe you're right. I'm thinking about building a SSD RAID0 for the manager (~2x500MB/s) or two separate SSDs, so it would be nice to know if this really makes a difference.

[polynurb]

the brutal thing is heavy read/write request to the same drive at the same time. even a single ssd is hit by this.
it appears to me the merging is a pretty linear addition of data not needing too much calculation to be done.

when i merge with network it is done on a raid 6 with areca controller, but that is not even close to the speed that i get when merging from ssd to ssd.

[numerobis]

ah ok, interesting... so then an option to set a different output directory for network merging would be nice...