Forum

Author Topic: Fine-level task subdivision performance  (Read 2695 times)

andyroo

  • Sr. Member
  • ****
  • Posts: 438
    • View Profile
Fine-level task subdivision performance
« on: May 01, 2020, 01:06:30 AM »
Has anyone benchmarked a good-sized project with fine-level task subdivision enabled vs disabled? Using 1.6.2 I'm just curious if it speeds up or slows down processing if you're otherwise not limited by RAM. Running on AMD 3960x 256GB RAM 2 2080 Super GPUs. Could try re-running same ~7k image project if nobody has an answer.

neavrys

  • Newbie
  • *
  • Posts: 18
    • View Profile
Re: Fine-level task subdivision performance
« Reply #1 on: May 05, 2020, 08:41:04 PM »
Do you find a difference?

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: Fine-level task subdivision performance
« Reply #2 on: May 07, 2020, 02:22:00 AM »
Hello andyroo,

Most likely the processing time with the fine level task subdivision will be longer, as the intermediate processing results are saved to the project folder first and then read, when they are required on later processing sub-step.
Best regards,
Alexey Pasumansky,
Agisoft LLC

andyroo

  • Sr. Member
  • ****
  • Posts: 438
    • View Profile
Re: Fine-level task subdivision performance
« Reply #3 on: May 12, 2020, 12:15:14 AM »
So I had interesting results with this. FLTS ON seems to give both FASTER and BETTER alignment results.

TLDR; Same set of 7148 JPGS got 6% more images aligned in 10% less time with Fine Level Task Subdivision turned on.

I was benchmarking DNG/JPG with FLTS OFF and 7148 images. DNG aligned 6909 (5:28:19) and JPG aligned 6550 (6:50:03), which bummed me out because DNG was better and faster (but 16-bit and waay bigger files). Interestingly a colleague who had the same set as me and a machine that ran slower than mine in other benchmarks ran the DNGs with FLTS ON and got comparable timing (5:28:01) and 6981 images aligned. I decided to try the JPG again (copied the chunk into the same project, reset alignment, ran with FLTS ON, and got 6925 images aligned (6:10:10)

jedfrechette

  • Full Member
  • ***
  • Posts: 130
  • Lidar Guys
    • View Profile
    • www.lidarguys.com
Re: Fine-level task subdivision performance
« Reply #4 on: May 12, 2020, 02:36:04 AM »
Depending on the dataset, alignment results seem to be pretty variable from run to run even when using exactly the same settings so I'd be a little wary about interpreting to much based on only a few runs (n=4 in your case?).
Jed

andyroo

  • Sr. Member
  • ****
  • Posts: 438
    • View Profile
Re: Fine-level task subdivision performance
« Reply #5 on: May 12, 2020, 10:16:38 PM »
Using a subset of the data I've been trying different DNG-JPG conversion (all with FLTS enabled) and getting within 1-2 images (out of a thousand) for over 30 runs, and within one image (out of 120) for over 40 runs for another subset. Haven't tried FLTS on/off for these smaller sets, but that's an interesting idea. Based on my experience with FLTS improving both alignment and speed though I don't think I'll be spending too much time exploring performance with it off.