Can't figure out why WebODM isn't utilizing all of the available RAM in my Google Cloud VM instance

[cmccurdy]

First, I’d like to thank those who have contributed to this project! I’m having a lot of fun learning how to use WebODM and get the most out of it.

Background: my company is starting a drone department, with a main goal of being able to map cell towers. I’m really only interested in generating a point cloud (no orthophotos, meshes, or DTMs/DSMs) from the drone images. We’ve demo’d Pix4D mapper & inspect, and while I really like the results and low processing time of Pix4Dinspect, I’m not sure it’s going to be cost-effective. I’ve been playing around with running WebODM in a Google Cloud instance, following the instructions at the end of the README file.

The issue: I was able to successfully process a few sets of images (200-300 images @ 20MP), however, upon opening the Diagnostic menu while the jobs were processing, I noticed that only a small portion of the RAM I provisioned was being used. I think it maxed out at 8 GB, out of 128 GB. I don’t have much experience with Linux, VMs, or Docker, so I’m kind of at a loss with how to troubleshoot this.

1). I set up a Google Cloud VM, using the “e2-highmem-16” machine type (16 vCPUs, 128 GB RAM), and attached a 100GB SSD volume.

2). I followed the instructions in the README file for installing WebODM on a Google Cloud VM instance

3). I uploaded my images and started processing

Things I checked/tried:

• set “max-concurrency” to 16 in WebODM web interface before starting processing
• ran “docker stats” to view memory usage/limits: memory usage hovered between 1-4 GB, peaking at around 8GB, and memory limits were 125.8 GB according to docker stats

Any ideas what could be limiting the amount of RAM being used?

Thanks in advance!

[smathermather]

Most processes aren’t RAM limited in OpenDroneMap. Only texturing spikes up the RAM. If you have a good balance of cores and RAM (4GB per core is the rule of thumb), you won’t see much available RAM used except at the texturing step.

[cmccurdy]

Oh, I see. Thanks for the quick response!

So if I want to speed things up, should I try using an instance that’s more CPU focused (more vCPUS)?

I’m really just interested in generating the point cloud.

[smathermather]

Oh yeah! If you aren’t generating textures, throw lots of CPU’s at it. I have no idea what the threshold should be for that, but MVE, the dense point cloud library we use, doesn’t require much ram and will max out whatever you have for your CPUs rather nicely.

[cmccurdy]

Sweet! Thank you so much!