(Fixed, Fixed) [HOU-2682, HOU-2604] rendering alembics is slow

[ronald_a]

Hi,

it appears, that rendering alembics is slower (about 2x) than rendering actual geometry. Is this to be expected? I am actually a bit surprised, because we have done so many projects using alembics but it seems I never compared the renderspeed to actual geometry. I am a bit suspicious that this might have another reason (system updates, etc.) so I created a simple testscene. If someone else could have a go at this and let me know if they see the same difference in renderspeed.

the difference seems to be less with cuda cpu and almost neglectable on cuda gpu. I am actually interested in cpu side of things, though.

Here is the file, give it a spin:


https://we.tl/t-SGAGPoRfhH

[ronald_a]

Digging deeper into that issue, it also seems that packed/unpacked alembics are looking differently due to what seems to be a difference in normals (the original alembic had normals). it would be great if someone could confirm this.

[bdancer]

On my 2950X:
RENDER_THIS: 0h 1m 14.3s
AND_RENDER_THIS: 0h 0m 51.9s

On QAs machine it's 3m vs 3m 30sec, so I really can't preproduce 2 times slower rendering...

[ronald_a]

Hi bdancer , thank you for looking into this. It seems that the more cores the machine has, the bigger the difference is. On lower core count machines (dell 2x 8 core xeons) we get the same difference as you do (and +40% longer rendertimes seem to be quite a lot as well). But on the higher end (custom 64 core threadripper, dell precision 2x20core xeons), it's a lot worse. I suspect the QAs machine was less than 16 cores.

So it might have something to do with high core count but not necessarily with numa (refering to your response in the alembic displacement thread).

[tsvetomira.girginova]

Dear ronald_a, hello.

Thank you for the clue and for reporting it!
I have tested it on even more powerful configuration and the issue is now reproduced and added to our system with HOU-2604 number.

[GoodbyeKansas]

Nice that you could reproduce it! This is maybe something we have been hitting as well lately when we have seen some renders take much longer than expected. And we have a gazillion alembics.

-- Erik

[ronald_a]

Dear ronald_a, hello.

Thank you for the clue and for reporting it!
I have tested it on even more powerful configuration and the issue is now reproduced and added to our system with HOU-2604 number.

tsvetomira.girginova Awesome! I would like to add that - allthough stated differently before - it seems that numa DOES influence the speed. The slowest are the machines with higher core count (40) and two cpus - single cpu machines with only one cpu (64 cores) seem to be a bit faster (relative to rendering unpacked alembics).

Did you also have a chance to look into the issue with the normals? My testscene gives a perfect example for that.

[tsvetomira.girginova]

Fido Film, great that it could be helpful for your work too.

ronald_a, good observation about the CPU count, will add this to the issue.

About the normals question, we saw a little difference in results, yes.
You will be notified when something new comes up.
Now it is on further investigation.
Thank you x2.

Best regards,
TS

[ronald_a]

Hi tsvetomira.girginova, I was wondering how things stand with this issue. I saw that this (or a very similar issue) has been fixed for vray for maya. It would be really great to get this fixed in houdini as well rather sooner than later because we see a lot of situation where this increases the rendertimes greatly (up to the point we have to work around it).

[tsvetomira.girginova]

Hello!
I am out of office this week.

Before that I have tested it on many powerful configurations with test scenes including alembics provided by you. Could approve that most of the cases are already handled. But there are still some where the issue is visible - the work is in progress, so don't worry.

I won't spare you the big YES.
More details - when I am back to work.

Sunny greetings!

[ronald_a]

@tsvetomira. Hi there! I was wondering what the status of these issues are? These are pretty important to us.

[tsvetomira.girginova]

Hello, ronald_a.

This topic opens a big discussion and interesting work to do. So, again thank you!

You are already able to see a difference in render times with many alembics, but regarding especially the mountain test scene provided by you - another issue is opened to successfully solve it till the end.
It's in the final stage now.

In conclusion - this is pretty important for us too.

As promised, I will keep you posted.

[ronald_a]

tsvetomira.girginova Excellent! Do your efforts also include the issues with the normals that I briefly mentioned?

[tsvetomira.girginova]

ronald_a, please confirm that this is the kind of difference that you are detecting too.

[ronald_a]

tsvetomira.girginova yes, exactly that. It doesn't seem like much, but if you don't subdivide the geo (from fractures for example) you can sometimes see pretty dramatic differences.