RealFlow 2014 doesn't use all threads in new workstations

[keleb78@o2.pl]

Hello!

I've just tested RF2014 on my new workstation based on dual Xeon 2699 v3 (36 cores, 72 threads). It uses only one CPU, second one sits idle. It's a common issue related AFAIK to Windows dividing more than 64 threads into groups - at least that's what I heard from Mark Granger from NewTek. It's solvable though as Cinebench R15 shows - it uses all 72 threads.

Posted it here, 'coz for some reason I can't get to RF Tech Support from inside Portal.

Keleb

[Zakkorn]

As Keleb said I have the same problem, if not, worst.

I own both an old i7 Nehalem (6.5 years old) and a new dual Xeon 2690 v3 (2 weeks).

The new is faster, core per core or thread per thread (whatever you prefer), in every benchmark I have test it. Including also 3ds max and Vray.
In full utilization of the threads is about 6 to 7 times faster.

The strange problem is that if the new uses all the threads (48 in total) in Realflow, its very much slower than the old with the 8 threads. We are talking about 5 to 10 times slower.
If I put both machines to use only 1 thread, the new one is about 5% to 10% slower.

Is not only that Realflow doesn't utilize the threads well, but at the same time is the only program that the dual system is slower than the 6 year old even core per core!!!

An I doing something wrong?

[LuisMiguel]

Hey,

Could you tell me how many particles do you have in the scene?.

luisM.

[Zakkorn]

The problem appears from the 1st frame. So doesn't matter if there are hundreds or thousands particles.
My test scene begins with lower than 200 particles.

[tsn]

Your particle counts are really small and so it's no wonder that RF doesn't use all available cores. It takes a some time to split the simulation and send the chunks to the cores. With just a few hundred or thousand particles this preparation step often takes longer than the actual simulation step. That's why RF doesn't use all cores/threads by default.

The fact that your new Xeon is slower than your old i7 doesn't necessarily mean that RF is doing a bad job. These i7 processors are really fast and normally outperform Xeons when it comes to single-threaded jobs. Even when you compare rather old i7 systems with new Xeons.

Did you every try to simulate a big scene with your new computer, e.g. a Hybrido scene with 20 or 30 million particles? This is the kind of simulation that will really show you the advantage of multi-core Xeons. CPU load should be between 95% and 100% most of the time. You'll also see performance decreasing when a frame is finished. Then, RF writes the data to disk and prepares the next frame. Furthermore, some processes in simulations are very multi-threading friendly, others are not, and then it's better to perform the step in a single thread rather than splitting it.

Another reason is that there are also OS and/or hardware bottlenecks, but in this case the list of possible problems is nearly endless.

I hope this gives you some insights

[Zakkorn]

No, I havent tried Hybrido yet.
So definitely I will give it a try.

Also after a few testings I found that waw my mistake that the Xeon showed slower than the old i7.
In fact its about 10% faster core to core.
And I test some lets say heavier scene and I noticed that the more the particles the more utilization of the cores.

We are talking about a 5% to 10% better speed with each core that is added (if this core would be used, though)*

And I came to a conclusion that with about 12 threads in 1 million particles you are doing better than with 10 threads or lower.

Its sounds complicated but thats the conclusion.

Thanks for your answer.

*If the core would not be used then will have a worst result in timings.
For example in a test I found that till about 1 million particles the 12 threads were giving better rendering times than the 16 threads.
But above 1 million the 16 threads began to show better results.
Thus if I had continued this test, at about 2 million probably I would have better results with 20 threads maybe and so on.