CPU and memory usage?

[TSR_trix]

I have a question about Phoenix´s way to use CPU´s and memory.

We have a computer for simulations with following specs:
Dell
Intel Xeon CPU E5-2450 0 @ 2.10GHz (32 CPUs)
64Gt RAM

I have made a liquid simulation with adaptive grid(limited to 90%memory).
Settings at the beginning are conservation quality 80, advection method classic, maximal step 6, upper 20, lower 1.
Cell size 0,081m, total cells 29 million.
Adaptive grid Temperature, threshold 1000

What actually happens during the simulation process?
I mean, we have monitoring CPU and memory usage during the last frames which take almost 8 hours to simulate (grid size is 630 million), and most of the time CPU usage is very low (3-7%) and memory usage is near 62 GB.
Im wondering why CPU usage is so low most of the time? Also it seems that only one core is doing anything at all for some moments.

[Ivaylo Katev]

in the liquid simulations the most CPU time is spent in the conservation process, and unfortunately only the smooth method is good for liquids. this method is serial by nature, but nevertheless is extremely fast. of course we have to rework this, it's weird to have only 3% CPU usage. in version 1.2 we had an option to run in fully parallel mode, but there were some problems and it was disabled.

[TSR_trix]

Ok. Thanks Ivo.
I´m pleased about the speed of the Phoenix simulation overall, but what disturbs me is the fact that I cant get actually any benefit from the multiple account of CPU´s (power of the machine).
If I run the simulation with my own machine (i-7 2600 3.4GHz 8 cores) and in our simulation machine, theres not too much difference between simulation times even when I have 26 cores less on my own.

I made some tests with simple smoke sim (smooth conversation) in our sim machine, and again the CPU usage pumps all the time between 2%-70%.
What I would want to see is the CPU usage be at least 70-100%. Otherwise I could have only one (?) powerful CPU and get the same result with speed than now with so many CPUs waste a performance.

I hope in the future we can get every piece of performance into our simulations!

ps. With buffered conversation mode the CPU usage is much more, like 80%.

[Ivaylo Katev]

yes, the buffered method is parallel by nature, but it can't be used for liquids