Mert Yildiz
Ph.D. Candidate in Information & Communication Technology
Sapienza University of Rome, Italy
I design scheduling policies for large-scale computing clusters — simple enough to deploy, strong enough to perform — and extend them to AI inference serving and connected-vehicle networks.
Bio
My research starts from a question that looks simple but sits at the heart of modern computing: how do you keep a large-scale cluster busy without keeping users waiting? I design scheduling policies that are simple enough to deploy in production, yet strong enough to approach optimal performance — and I validate them against real workloads from Google Borg and Alibaba. This line of work has recently extended to AI inference serving and connected-vehicle networks.
Research Interests
Cluster Scheduling & Resource Allocation
Designing dispatching and scheduling policies for data centers and HPC systems that keep machines busy while minimizing user waiting times, validated on Google Borg and Alibaba production workloads.
Energy–Delay Trade-offs
Jointly optimizing a cluster's energy efficiency and response time in multiserver systems — without requiring prior knowledge of job sizes.
Job Runtime Prediction
Machine-learning models that predict how long HPC jobs will take, so the scheduler can place them well before they start.
LLM Inference Serving
How multiple large language models share a single GPU server — offloading, preemption, and cache migration to keep latency low for all of them.
Vehicular Networks & 6G
The timeliness of safety-critical information in V2X communication — Age of Information and delay metrics for connected vehicles as networks evolve toward 6G.
Selected Publications
- . “Towards Multi-Model LLM Schedulers: Empirical Insights into Offloading and Preemption.” 2026 Mediterranean Artificial Intelligence and Networking Conference (MAIN), 2026.
- . “Dispatching Policies in Data Center Clusters: Insights from Google and Alibaba Workloads.” Performance Evaluation 172, 102551, 2026.
- . “From Latency to Value of Information: A Review of Timeliness Metrics for Safe Transportation Systems in the 6G Era.” IEEE Vehicular Technology Magazine, vol. 21, pp. 41–50, 2026.
- . “Data-Driven Workload Generation Based on Google Data Center Measurements.” 25th International Conference on High Performance Switching and Routing (HPSR), 2024.
Experience
Doctoral Researcher in ICT
Sapienza University of Rome, Italy — performance analysis and scheduling for large-scale computing systems.
Visiting Researcher
IMDEA Networks Institute, Leganés, Spain — supervisor: Prof. Marco Ajmone Marsan.
Operations Analyst → Lead Ad Operations Analyst
Compulse (formerly ZypTv) — automated reporting pipelines, cross-platform ad analytics at scale, and leading a team of operations analysts.
Contact
I am always open to academic collaborations and interesting conversations about performance optimization, distributed systems, and LLM infrastructure.