Displacement slows down DR slave...

[iTiunov]

After enabling Displacement in Global Options the speed of local rendering goes down twice for example but speed of DR slave goes down 200 times. There is no dependence with bucket size. Network load is far from 10%. What's this? Thanks.

Take a look like slave (ex) buckets late for local (st) buckets:

[iTiunov]

Sorry. It's not cause of DR. Take a look:

We have:
1) Mac OS X 10.6.7, One Intel Core i7 @ 2.8 GHz (8 active cores ≈ 22.4 GHz totally), 16 GB of RAM;
2) Windows Server 2008 R2 SP1, Four Intel Xeon E4870 @ 2.4 GHz (64 active cores ≈ 153.6 GHz totally), 64 GB of RAM;

Render times are:
With displacement:
1) ≈ 17 minutes;
2) 2% done for ≈1 hour, canceled;

Without displacement:
1) ≈ 7 minutes;
2) ≈ 53 seconds;

P.S. Maya 2012 Hotfix 1, V-Ray for Maya 2.00.04...
P.P.S. The same simple scene with same settings. Scene is located locally...
P.P.P.S. Dongle is attached to Macintosh.

[iTiunov]

3) Windows Server 2008 R2 SP1, Four Intel Xeon E4870 @ 2.4 GHz (8 active cores ≈ 19.2 GHz totally), 64 GB of RAM;
4) Windows Server 2008 R2 SP1, Four Intel Xeon E4870 @ 2.4 GHz (16 active cores ≈ 38.4 GHz totally), 64 GB of RAM;
5) Windows Server 2008 R2 SP1, Four Intel Xeon E4870 @ 2.4 GHz (32 active cores ≈ 76.8 GHz totally), 64 GB of RAM;

Render times are:
With displacement:
3) ≈ 5 minutes and 15 seconds;
4) ≈ 25 minutes and 15 seconds;
5) ≈ 1 hour 9 minutes and 1 second;

Without displacement:
3) ≈ 2 minutes and 3 seconds;
4) ≈ 1 minute and 35 seconds;
5) ≈ 56 seconds;

P.S. Very strange...
P.S. Bucket size is 64*64 pixels...
P.P.S. Number of cores multiplied on core speed is not exactly equal to real CPU speed cause of Intel TurboBoost technology...

[iTiunov]

Do V-Ray manually manages threads or uses something like OpenMP? I mean is it an operation system bug or feature of V-Ray?

[Zdravko]

Could you please check if you are using identical V-Ray for Maya version on all machines?

[iTiunov]

Latest version downloaded from website. BTW the latest test are not DR. They were run on single machine.

[iTiunov]

The same regularity traced on 4x AMD 6197 machine... But it slows down less that 4x Intel 4870... I'll put render times here later today.

[iTiunov]

e-ex stands for Intel 4870
n-g4 stands for AMD 6176
and there is no lhost (localhost) buckets. they are done for this moment.

BTW there is not DR problem. It's just a quick way to visualize speed difference between 8 cores and more that 32 (48 AMD and 64 Intel) cores...

[iTiunov]

Here they are (displacement is on):

Intel - 8 threads - 0h 05m 31.1s
Intel - 16 threads - 0h 25m 21.9s
Intel - 32 threads - 1h 11m 4.3s
Intel - 64 threads - too long

AMD - 8 threads - 0h 6m 28.3s
AMD - 16 threads - 0h 5m 13.1s
AMD - 32 threads - 1h 40m 21.3s
AMD - 48 threads - too long

P.S. Intel - 4x 4870 with 64 GB RAM (64 logical cores, HyperTreading is on).
P.P.S. AMD - 4x 6179 with 64 GB RAM (48 physical cores).
P.P.P.S. Not absolute values are interesting but speed changes depending on number of threads.
P.P.P.P.S. All render times are averages of 3 attempts.
P.P.P.P.P.S. I'll do some test with HyperTreading off on Intel later today.

[Zdravko]

What value is your Dynamic memory limit set to?
Does the problem persist without the DOF effect?

[iTiunov]

The Dynamic Memory Limit is set to 16384 MB. Actual memory allocation is about 8 GB. I'll try to turn off DOF today later. Thank You for reply.

[Zdravko]

Set the Dynamic memory limit to 7500 and render again.

[vlado]

Can you get me that scene to vlado@chaosgroup.com? I've tested this, but I can't reproduce the slowdown here on machines with 24 logical cores (both AMD and Intel).

Best regards,
Vlado