How can I optimize this FLIP sim

[art3mis]

Hi

First of all a HUGE thanks to everyone in this great community. Pretty well impossible to master Houdini without the support here.

In an effort to try and optimize a fairly simple FLIP simulation I have been re viewing as many tutorials as possible. In particular this one on Houdini 15 - FLIP Workflow Enhancements HD by John Lynch

What has me 2nd guessing my particular setup is at 7:31 in the video John suggests you should be able to scrub through your simulation when you have the compressed_cache node view enabled.

I cannot. Huge bottleneck and all I see is Cooking and Solving messages for several minutes whenever I switch frames on the timeline.

My machine specs are

i7 6900K CPU

TitanX Pascal GPU

32 GB RAM (upgrading to 64 is me next step)

My particular FLIP flat tank simulation has a particle Separation of .04 with approx 15 million total points or particles. So I am wondering if THAT is the bottleneck and I should be targeting a much smaller total particle count?

Anyways was hoping someone with more FLIP experience than myself could have a quick look at the attached .hip and offer any tips or suggestions to better optimize it.

 

 

helloWorld_020_LARGE_.hiplc

Edited November 10, 2016 by art3mis

[Atom]

If you take a look at the memory footprint on your scene it is over 50% larger than what John is working with in the video demo. His uncompressed memory is about 256MB and your scene is 600Mb.

I took the time to bake out a few frames of your sim, following John's instructions from the video, and I was able to achieve similar results that he describes. A fairly light weight, scrubbable, display preview. I did make a few tweaks to your settings because I am on a low end machine (4 core AMD 3Ghz with 14GB ram). I set the particle separation to 0.1 and set the feedback scale to 1 and connected only one letter.

 

Flip really is one of those beasts that requires compromise. If you set your particle separation too low you just end up bottle necking everything and waiting forever. That is why you see a lot of motion blur used in "high-end" fast action fluid simulations. They're blurring the fact that the simulation is not really that hi-res.

 

 

Edited November 10, 2016 by Atom

[art3mis]

Thanks Atom,

Based on further tests (and rewatching John Lynch's excellent video a 3rd time) I would have to say though that I cannot reliably change particle separation and get consistent results.

Also in my case I want the objects(letters) to sink to the bottom of the tank which they won't with a Feedback of 1.0 (unless I were to crank up their density to a very high value)

My feeling is also that THE 2 most important things when using FLIP tank shelf tool is to get the scale (size) of the tank and associated Particle Separation correct and to NEVER change these when making further changes. In John's demo he is using a particle Separation of .005.

If you try and reduce the particle count to say .05 his demo essentially 'breaks'. There are not enough particles to get any sort of surface detail happening.

The FLIP tank I was using was WAY too big giving me an unmanageable number of particles at the necessary particle resolution.

Also learned in John's video the importance of using adjustable bounds and camera frustrum option in the surface preview nodes to get quicker previews without changing the dynamics of the simulation.

For me it seems the learning process for FLIP is never ending...