Strange workstations testing results.

[tincansassoc]

Can you provide a set of images for everyone to run? Your benchmark comparison should be more accurate if everyone runs the same data sets. Maybe two different example sets and specify the settings we should try tweaking. If bandwidth is an issue maybe you could upload the datasets as a torrent file and we can do P2P transfers. Just a thought.

[Wishgranter]

Yes it will be something in this way, 2 dataset for sure, one smaller 32 images and one bigger 50-70 images with proper image size so its not take hours to finish :-)

because the difference is going to be higher as number of images is higher..... morning will prepare it.....

[glennn]

I am playing with a new Xeon too and noticed the speeds were dismal and found this thread.
I noticed my home system was considerably faster and that I needed to correct some settings to bring this Xeon workstation up to a reasonable speed.
The sample01.zip scene seems to run at a decent speed.  Its only when processing my own image set where it is a huge difference in processing speed
Image resolution of the sample images are 2184 x 1456 @ 32 images
My personal images are 5184 x 3456 @ 45 images

Xeon system - Dell Precision T3610 -
Windows 7 64-bit Enterprise
Intel Xeon CPU E5-1650 v2 @ 3.5GHz
32GB Ram 1866MHz
Quadro K4000 - 3GB

I changed settings to use 11/12 cores. Is this still correct practice?

Testing the sample01.zip scene on Xeon
Note : This seems to run at an ok speed.

align photos : High points 40,000 > 130.748 sec
build dense cloud : high > 248.198 sec
build mesh (dense cloud) 1,000,000 > 63.719 sec
build texture n/a
------------------------------------------------------------------------------------------
Testing my own scene on Xeon

align photos : High points 40,000 > 296.753 sec
build dense cloud : high > 4790 sec  (took about 2 hours ?)
build mesh (dense cloud) 725sec
build texture 89 seconds

I will post the differences between the Xeon and my i7 from home.

[Alexey Pasumansky]

Hello Glen,

The difference in dense cloud generation stage is understandable, since the size of your dataset is 1.5 bigger by the number of images and the resolution is almost 5.5 times higher. So in terms of resolution, using High for your dataset will be almost equal to the use of Ultra for sample01 data.

Also please note that in case you wish to test only CPU performance for both configurations you need to uncheck OpenCL devices (I believe that at home you should have something faster than Quadro4000).

[VoRo]

I can confirm an unexpected performance on a multi-processor system
(4 x Xeon E5-4640 + Nvidia Tesla k20c + Nvidia Quadro k5200, 32 cores/64 threads, 30/60 used).
For comparison I used the recommended building dataset with 50 images in medium resolution.
The runtime is really disappointing. However, other application specific benchmarks on this server show quite promising results (see table attached)

[Wishgranter]

VoRo can contact me on muzeumhb@gmail.com propably have some not optimal settings on it...