Computing -

[traiano]

Hi.

I want to know something about Houdini Rop Cache and Computing.

Im simulating 240 Frames of dust with smokesolver emitted from particles.

Pyro Divisions are 0.06 - Volume source scale divisions: 0.1

Im caching and from Frame 0 to Frame 100 it was only 4 hours.. .bgeos weight from 18 Mb at early start to 200 Mb each bgeo file.

So now the caching and is taking ages.. It took 17 hours to cache from Frame 100 to Frame 133.

Bgeos are getting as big as 450mb each.

How long it will take to cache until Frame 240?

Is It normal that a bgeo weight 400mb and take 1.5 hours to compute each frame?

Im using Open CL - Quadro 4000

Edited March 7, 2013 by traiano

[Hudson]

Hi traiano!

Since I am new to houdini I can't answer you if it is usual for a frame to take that long. But first of all I would check if you have scaled everything correctly. Expect a meter to be one unit as a default in houdini. If you know the irl scale of your sim try to match it since I think a scene to large scaled can require a bit of your sistem.

To answer your question I simply used mathematics. Using a quadratic exponential function I approximate your total sim to take 70 hours so you have two or more days to go. Well as I said it is only a mathematic approximation. I hope any houdini masters could help you since their experience is more valuable than maths.

If you want to make an even better approx create a simple graph in Xcel with the different frame values and also the time your sim has been running or even better put that in matlab. A bit overkill I would say.

Here is just a simple graph from excel

Hope this gives you a hum of sim time. Would be cool if you could report back ans see how far I actually was from the right answer

[Filipp Elizarov]

HI Traiano.First of all we'd like to know which fields you've cached ?Usually a lot of people trying to cache all fields from pyro object - like temperature , heat , fuel , rest , rest2 , vel. But for dust you need to export only density or if you'd like to advect particles for more natural looks after sim you shoul ad 1 more field VEL . Export only density and vel.You don't need to use temperature after sim at shading time or heat or rest(rest fields are vectors and they will have a huge size).And for a bit smaller size cache your sim to name.bgeo.gz - gz will help to archive each bgeo file and compress its size very well.And don't use open CL for big simulations because it won't help you so much and can slow down sim process.

[tjeeds]

You didn't turn Initialize Simulation Ops on by any chance, did you?

[Nerox]

It sounds more like your sim fits in the gpu memory for the first part (assuming the box grows) and then switches to the CPU. I would advise to run it on the CPU entirely.

Depending on your voxel resolution and gpu memory, you can only fit a given number of voxels. For the GPU to work it needs to store all the data in the GPU memory, if it can't it will fail. Instead of failing the sim Houdini switches to the CPU, how ever this can cause artifacts.

GPU is great for previewing, but most of the time you want to sim you final sim on the CPU (non-opencl)

[traiano]

Hey guys thanks for your input.

NIck, indeed Houdini crashed at Frame 137 after 38 hours of caching time.

So next time Im using OpenCL for mid rez preview, then CPU for the final crank up on divisions like under 0.09

So i have 138 .Bgeos some of them weight 450Mb.

What Can I do with these?

Is there any possibility to Cache from Frame 138 to Frame 240, and immediatly start computing at Frame 138 so I save my time Re Computing F0 to F138 again?

[traiano]

This is the 18th time I try to cache this dust simulation and houdini crash or stalls.. Som im very tired and I have a modern computer how come I struggle to simulate this basic smoke ?

Sometimes I think I should go back to 3ds MAx and FumeFx for this

[nMolson]

Hey Martin, is the container massive by the end? and are you clustering it?

Iv had tons of sims get way to big and slow to a crawl before crashing completely, Clustering it into multiple sims usually does the trick. Seems to be the only option for high res large scale pyro

[traiano]

Hey Nick!

I think clustering will be great, but I dont have any idea how to use it.. any link that explain it or can u tell me how to?

thanks mate

[nMolson]

http://www.sidefx.co...dyno/clustering

If your smoke is sourced from particles the entire time you can also just break up the cached particles into multiple objects before converting them into a source volume by deleting sections out with a bounding box and then run each separate particle cache into the sim one at a time then comp them together(possibly even render the caches together). If that makes any sense

Edited March 8, 2013 by nMolson

[traiano]

Yes, that would perfectly work! I will check it out.

thanks Nick

[traiano]

Hi Nick,

Its 5 to 6am and I have a new crash at F 180. I was trying 0.25 divisions this time as I lost hope to sim this under 0.09

I tryed to cluster but I couldnt do it, nothing happens when I play it.

Unfortunately there is no time to learn this now. Im working vs time.

I was collidinng dust with geo and now i removed this feature and I still crash at 0.25 divisions which is a crap detail box. So Im going to try this please tell me if it will work.

i will do an explicit cache from F0 to F100 at 0.09 divisions with OpenCl. this will take around 4 hours.

and then im going to load this sim, turn off Open Cl and explicit cache from F 100 to F 150 so that the video card does not complain and crash. Then I will do the same from F 150 to F 200

Do u think if i go with 0.09 divisions to .25 and then .35 this can affect my enire simulation detail flow to the point first cache doesnt match with the second and the third?

[traiano]

Well I decided to stay with same Divisions for the whole frame range.

Im explicit caching since Frame 90 without OpenCL and now Im on Frame 133.

Each frame is taking 32 minutes. Everything looks stable as Im not using the computer while caching.

I hope this way holds safe until frame 240.