Thelio Massive at the Lab: An interview with Luca Della Santina
Every now and then we like to check in on our customers to find out about what coolness they’re up to. This week, we sat down with Luca Della Santina, an assistant professor at UCSF in the Department of Ophthalmology, to see what he and his Thelio Massives are discovering at the lab.
What kind of work goes on in the Department of Ophthalmology?
Everything we do is focused on the eye and on vision. I am also part of the Bakar Institute, which is a computational institute specializing in machine learning and deep learning applied to health sciences. The lab that I run is divided between working on computational approaches, mainly image analysis.
What projects are you working on right now?
One major current project is detecting an infection of the eye called trachoma. Trachoma is an infection that affects the inside of the eyelid. It usually occurs in countries below the tropics, and it’s a major cause of blindness for people across the world—except for in wealthy countries like the US where it’s very rare. Eliminating it elsewhere is a major goal of the World Health Organization. Africa, South America, Asia and Oceania still have many cases, so we’re taking photographs of the afflicted eyelid to look at the sites where bacteria has infected the eye. Then we use deep learning to detect it automatically to help public health experts decide which communities will require antibiotic treatment.
We’re also taking images of neurons in the eyes and map the connection between them, called synapses, to study how degenerative diseases of the eye such as glaucoma can alter the wires between neurons. Knowing which neurons are the most susceptible to disease will shine a light on new and more sensitive tests to catch these blinding diseases before they can actually cause major vision loss. This type of research generates really large data sets, in which each image is large many gigabytes and for which the analysis is very computationally intensive, both for the GPU and the CPU.
How long have you been using System76 workstations for your projects?
We started to use System76 systems two years ago, give or take. It was part of setting up my computational lab. One of the goals was to have a completely open a stack, and your workstations were an integral part of this strategy.
What is the computational stack you’re using?
We have the Thelio Massives configured for deep learning and for processing large image data. One of the systems has NVIDIA Quadro RTX 8000 GPUs for training larger models than we usually do. In the other system, we have it configured with dual CPUs and dual NVIDIA GeForce RTX 2080 Tis. The reason for that is that some of the computational work is being developed with parallel computing, both on GPUs and CPUs. The more cores and the more CPUs we get this on, the better.
How do you balance workloads between the CPUs and GPUs?
Strictly for the projects I’m on, they’re each about as important. All of the machine learning runs off the GPU right now, but all of the basic image analysis and parallel computing actually works off the CPU. The reason for the latter is there’s no significant advantage to push that work onto a GPU. There are a few algorithms that we cannot parallelize on the GPU because of the way they are designed, and one of these is actually pretty fundamental in the way we segment images, so if we put it on the GPU there is not much increase in speed because we cannot push it onto every core of the GPU. For most of it, we need the raw power of the CPU.
What were the determining factors when you decided to go with System76 and our Thelio Massives?
A few things. We wanted a system that was designed to run Linux from its foundations. There are not a lot of systems like yours, so that was a major factor in our choice. We also wanted a system that we could expand easily in the future, and we found out that the Thelio Massive has has great expandability.
The most important factor for me was being able to double or triple the RAM somewhere down the line, and maybe have another couple of GPUs in the system. Having storage options is useful for us because we may generate a dataset and on a single 4TB hard drive, so the ability to just pop out and pop in hard drives is very easy. It’s actually huge for us. I ended up buying a bunch of 5TB drives and just packed them in. Most of the small stuff we just run off of the NVMe drive, and that’s much better than the rest of the storage we have.
I really enjoy how quiet these machines are! I can testify that we’re sharing the same room with another computer from a different vendor with similar components, and it’s about 10 times louder than the Thelio Massives.
What operating system do you use?
So far we’ve been keeping both Thelio Massives on Pop!_OS. The other workstation we have in the lab is either Ubuntu or Windows.
How has Pop!_OS been for you?
The software pipeline we use runs out of the box pretty well on Pop!_OS, so that’s not been an issue so far. I appreciate that you guys have full disk encryption out of the box.
We’ve also heard you’re thinking about buying a Lemur Pro. What made you consider that machine?
I need something that’s light that I can bring around with me. It’s also got a good number of ports, which lately has been hard to find on a laptop, which frees me up from having to carry dongles on my trips. I can also configure it up to 40GB of RAM, and I need at least 32GB, so that’s perfect for me.
Would you like to share how System76 has improved workflow for you and your organization? Contact [email protected] to set up an interview!