PhD#2: Ehsan Yousefzadeh-Asl-Miandoab

 

Note: On April 9, 2025, my second PhD student, Ehsan Yousefzadeh-Asl-Miandoab, successfully defended his thesis (here is the preprint). I said some words about him after the defense. The text below was my write-up to be prepared for those words.

 

Ehsan started his PhD three months after Ties, but I knew of Ehsan before Ties.

By the time Ehsan applied for my PhD opening, I had already assessed Ehsan twice for different openings in our group. The first one was for Zsolt István’s PhD position and the second was for the DAPHNE project, both for very hardware-oriented topics that included FPGA programming. Ehsan was one of the strongest candidates for those positions as well, but world politics interfered. While FPGAs are the trickier hardware platform, when it comes to the US sanctions against Iran, the rules for GPUs are more relaxed, so I was able to hire Ehsan for my project.

 

Ehsan’s start at ITU completed the founding members of the RAD group: Ties, Ehsan, and I. Us three had a brand new group to establish, so came the website and the logo, where Ehsan led the efforts. 

 

In the beginning, I had to convince Ehsan to pace himself, and that is a challenge for someone like me who doesn’t know how to pace herself before experiencing exhaustion signs. Ehsan wanted to do his research, maintain an excellent github repo for all the code and data he creates, prepare extensive teaching material, record youtube videos for that material, write technical blog posts about the papers he has read, and fly rockets to the moon all at the same time. It is great to see that fire in someone, and one shouldn’t kill it. But doing everything in parallel isn’t possible, and one has to learn how to prioritize and adjust over time. I can’t claim I always guided him perfectly when it comes to prioritizing, but we established a more reasonable pace. 

 

For the core part his PhD, Ehsan took on a challenge where neither of us were experts on; using machine learning (ML) for systems. More specifically, he wanted to use ML to estimate memory needs of deep learning training workloads to effectively guide resource management and collocation decisions for these workloads. While testing different GPU collocation ideas, we found out-of-memory errors to be one of the biggest challenges against collocation and often overlooked in literature. Hence, we needed something to minimize them. We tried the more analytical techniques found in prior published work, but they led to drastic misestimations. So, Ehsan wanted to try out something learning-based to see if it could lead to more precise estimations. But, once again, we were not experts in ML. We tried and failed, tried and failed, on and on. Nevertheless, Ehsan persisted, kept asking for help from different ML experts, and got many good suggestions, but they did not lead anywhere either. In the end, Ehsan found the collaborator he was searching for all along; Reza. I cannot claim any credit in that search effort, it was all Ehsan, I just did my best to prevent him from giving up. With the help of Reza, Ehsan built his estimator, and the last two chapters of his thesis got unblocked. 

 

Ehsan also persisted despite the many challenging news coming from his homeland and not being back home since he moved to Denmark for his PhD. While I mention this here as part of my praise of Ehsan, I want to emphasize that no one should go through this, especially to get any sort of praise. 

As a result of this, some of our one-on-ones with Ehsan turned into venting sessions commiserating over politics and PD3.

 

And finally, a couple of fun facts. 

Ehsan probably knows Turkish music better than I do. I don’t know how many times I heard Turkish tunes coming from his office or headphones.

Early in his PhD, Ehsan decided to get some plants for his office and got me one as well. As someone who never managed to keep plants alive at home, including cacti, I somehow still manage to keep this plant alive, while Ehsan killed all of his. I even had to smush the alive parts of the plant back into the soil after its root broke off from the rest, and it keeps flowering, as you can see below. Looking at it is a good way to remind myself to keep being stubborn, and I hope I keep motivating my students to do the same.

 

Congrats for a well-earned PhD! Looking forward to collaborating further with you.

 

February 25, 2025 - after the root broke off from the rest of the plant.

April 10, 2025 - the current state.