Talks and presentations
⬗ Resource-Efficient Training via Task Collocation on GPUs
Workshop: D3A 1.0. | February 1-2, 2024 | Hotel Nyborg Strand
– The #presentation was prepared for the Resource-aware machine learning workshop. Also, I presented my PhD work in the posters hall..
⬗ Profiling and Monitoring Deep Learning Training Tasks
Workshop: The 3rd Workshop on Machine Learning and Systems (EuroMLSys) - May 2023
– This #presentation was prepared for the accepted paper to the EuroMLSys workshop. The #poster version of the presentation can be accessed here.
⬗ Orchestration of Deep Learning Tasks on CPU-GPU Co-Processors for Multi-Tenant Settings
Workshop: EuroSys Doctoral Workshop 2023 (EuroDW '23) - May 2023
– The #presentation and the #poster were prepared for the accepted paper to the EuroDW workshop.
⬗ Overprovisioning GPUs in the age of AI
Courses: Large Scale Data Analysis - November 2022
– This #lecture was prepared for computer science master students at the ITU. It reviews machine learning, deep learning and GPUs as the primary general-purpose processors for deep learning training tasks, then the underuitlizatoin challenge of GPUs is discussed and three potential solutions of workload collocation for Nvidia GPUs (multi-stream, MPS, MIG) are discussed.
⬗ Processors
Courses: Large Scale Data Analysis - April 2022/ Big Data Management - November 2022
– This #lecture was prepared for data science bachelor students at the ITU. It reviews computing, processor, transformation hierarchy, tradeoffs of different processors, CPU (simple in-order processor, cache hierarchy, pipelining), GPU, FPGA, ASIC accelerator. A recorded version of the talk can be accessed here.
⬗ GPU Efficiency through Intelligent Collocation
6th International Summer School on Resource-Aware Machine Learning (REAML 2022) - September 2022
This #poster was prepared for presenting the undergoing work in the students' corner of the TU Dortmund's summer school. It discusses the underutilization problem of GPUs, especially for deep learning training tasks, the causes of the problem, available collocation options and their performance comparison, and finally next steps.