CPU/RAM for processing 2DFull - 2677 Images

Really new to OpenDroneMap. I’m looking for recommendations on the ideal CPU/RAM needed to process 2677 images 2DFull. I have one running now with 96 vCPU/384 GB RAM, the process is still running and at 19 hours so far. I ran a fast-ortho same specs it took 21 hours to complete. With both options I’m looking for ways to get the process down faster. Should I increase CPU or RAM or both?

1 Like

Hmm… What speed is the vCPU? Some parts of our pipeline benefit from high clockrate more than core count since they are single-threaded (though possibly spawned in parallel).

Hi Saijin,
3.3ghz

1 Like

That’s pretty quick!

I assume you’re not getting any Exit Code 1 errors, right? So no Out-Of-Memory errors yet?

Maybe try upping the vCPU speed first (if possible, not sure what your host offers), and if not, slightly increasing vCPU count.

The run with 96 vCPU/384 GB RAM was my first successful run. But took 30 hours :slight_smile: . I tried a z1d.12xl using the following options too:

–dem-resolution 1.0
–dsm
–dtm
–feature-quality high
–mesh-octree-depth 12
–mesh-resolution 300000
–orthophoto-resolution 1.0
–pc-geometric
–pc-quality high
–use-3dmesh

That took 29:33 to process. A z1d.12xl has 4.0ghz, 48 vCPUs and 384GB of memory. Now I’m thinking if I run in a cluster and split, parallelize if I can Splitting Large Datasets — OpenDroneMap 2.6.3 documentation. Any thoughts here? I’m trying hard to at least get this processing at least under a day. Drone2Map took ~18 with a g4dn.12xl (4GPU, 48 vCPUs, 192GB RAM).

1 Like

Split-merge should help at lot up front, but you will have a slowdown at the end when all the work needs to be composited on one machine…

Hmm… Could be worth quantifying for your datasets

A couple of points of reference-

2214 images with resizing to 1/3 linear size, took 7h 41m with 96GB RAM, i7 8 core 3.8 GHz Intel processor.
It never uses more than about half the RAM though.

[Nombi - 15/09/2021] 2214 07:41:52

Created on: 20/09/2021, 12:32:48

Processing Node: node-odm-1 (manual)

Options:** dsm: true, dtm: true, optimize-disk-space: true, pc-quality: high, pc-rectify: true

1194 images took 22:47:54, although due to large elevation variations, there were gaps in the ortho due to not enough overlap on higher ground

Processing Node: node-odm-1 (manual)

Options: dem-resolution: 2.0, dsm: true, dtm: true, feature-quality: ultra, min-num-features: 10000, optimize-disk-space: true, orthophoto-resolution: 2.0, pc-quality: high

I’m trying to reproduce Drone2Map 2d full template run with the same data set to possibly save cost when running on AWS. My thoughts are to use less expensive, non-GPU intensive instances and get the job done in less time than with Drone2map. The 2D full template with Drone2Map output (per their documentation):

The options I’m currently using are:
–dem-resolution 1.0
–dsm
–dtm
–feature-quality high
–mesh-octree-depth 12
–mesh-resolution 300000
–orthophoto-resolution 1.0
–pc-geometric
–pc-quality high
–use-3dmesh

I may also need to play around with the options to match that of a 2d full with Drone2Map.

1 Like