https://dasya.itu.dk/semester/fall-2025/ Recent content in Fall 2025 on DASYA Hugo -- gohugo.io en-us © DASYA, IT University of Copenhagen Tue, 30 Sep 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-nas/ Tue, 30 Sep 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-nas/ Designing neural networks that are both accurate and efficient has become an urgent challenge as the cost of deploying machine learning systems continues to grow. Traditional Neural Architecture Search (NAS) methods usually optimize for accuracy under constraints such as FLOPs, model size, or latency. However, in practice the lifecycle cost of a model is not only determined by training, but also by inference. While training is a one-off expense, inference is repeated continuously once a model is deployed, and depending on the number of queries served, it can become the dominant contributor to energy use and cost. https://dasya.itu.dk/for-students/proposals/archive/pito-ml-orion/ Thu, 24 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-ml-orion/ This project investigates how running multiple deep learning training tasks simultaneously (collocation) affects performance on resource-constrained edge devices, specifically the NVIDIA Jetson Orion Nano. Students will deploy and benchmark various models (e.g., CNNs, Transformers) in isolated vs. different collocated scenarios, measure metrics such as GPU utilization, memory usage, training time, and analyze the trade-offs between throughput, fairness, and system efficiency. This project would be suitable as a BSc or MSc thesis. https://dasya.itu.dk/for-students/proposals/archive/pito-xnvme-duckdb/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-xnvme-duckdb/ A recent MSc thesis project at ITU integrated xNVMe to DuckDB. DuckDB, like many popular database systems, rely on a syncronous IO backend and the POSIX filesystem interfaces for its ease of use. xNVMe, in contrast, allows a unified interface for alternative and more performant IO backends such as io-uring and spdk. Thus, integration of xNVMe into DuckDB offers interesting avenues for data analytics workloads run on DuckDB. The existing integration of xNVMe to DuckDB is preliminary. https://dasya.itu.dk/for-students/proposals/archive/pito-edge-benchmarking/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-edge-benchmarking/ The work on running data-intensive applications on very powerful, expensive, and power-hungry server hardware is very popular thanks to the growing size of data centers and high-performance computing (HPC) platforms. However, with the rise of new generation internet of things (IoT) applications, the lower-power and lower-budget hardware devices that specifically target IoT, the edge platforms, embedded devices, etc. have also become quite powerful today offering non-negligible data processing capability. For example, NVIDIA Jetsons and Coral devices allow GPU- and TPU-acceleration of AI applications at the edge and come with hardware resources that can match the resources of a laptop. https://dasya.itu.dk/for-students/proposals/archive/pito-data-attribution/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-data-attribution/ Deep convolutional networks are able to learn representation of images, scoring well in tasks such as image classification and object detection. During model training, these networks have the ability to process different input sizes without requiring changes to their architecture. In this project, we would like to investigate the effects that changing input sizes has on these kinds of models. We want to do this by investigating and using data attribution methods, as these techniques can help explain the effect of input data on the predictions of the model. https://dasya.itu.dk/for-students/proposals/archive/pito-data-selection/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-data-selection/ Today’s foundation models are trained on vast amounts of data. The quality and size of this data has a huge impact on the accuracy of these models. Selecting the right amount and variety of data for a given task, however, is a resource-intensive process. In this project, which is part of a larger collaboration, we would like to expand our investigation of state-of-the-art data selection mechanisms from a hardware requirements and resource-efficiency perspective in addition to their impact on model accuracy. https://dasya.itu.dk/for-students/proposals/archive/pito-radt/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-radt/ Observing how well machine learning systems utilize hardware resources is a crucial preliminary step to improve system performance and reduce hardware waste. To do such observations, one has to collect a lot of monitoring data on hardware behavior through experiments. In our group, we have recently built a framework to aid the management of such monitoring data efficiently, called Resource-Aware Data Systems Tracker. There is a variety of work to be done on this framework to improve it. https://dasya.itu.dk/for-students/proposals/archive/pito-gpu-db/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-gpu-db/ This rise of hardware accelerators to meet the demand of AI workloads has also led to a variety of novel methods to leverage GPUs for traditional data analytics workloads. A key concern for any data-intensive system using GPUs is the efficiency of moving the data to the accelerator. In this project, we will investigate ways to improve data movement to GPUs by focusing on the steps of the data path and performance of different data movement options from the network or SSDs to GPUs such as GPUDirect or BaM [1]. https://dasya.itu.dk/for-students/proposals/archive/pito-tiny-resourcemanagement/ Tue, 15 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-tiny-resourcemanagement/ Today many data sources are small low-powered and hardware-constrained devices such as mobile phones, wearable or self-driving smart platforms, etc. Edge computing is a broad term that refers to computations performed on such edge devices. It becomes increasingly important to enable techniques that get more value out of data at the edge rather than always sending the data to a remote and more powerful hardware device in the cloud for further data processing and training machine learning models. https://dasya.itu.dk/for-students/proposals/archive/mh-generated-dataset/ Wed, 02 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/mh-generated-dataset/ (MSc Research Project / MSc Thesis) The goal of this project is to research how AI and large language models generate datasets. Research questions include: Where does the generated data come from? Are sources available on the internet or can they be found? What biases exist in the generated data? And how much of the data is simply wrong? Generated datasets are used in many fields in practice, including journalism (e. https://dasya.itu.dk/for-students/proposals/archive/mh-rbac-data-lakes/ Wed, 02 Jul 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/mh-rbac-data-lakes/ (MSc Research Project / MSc Thesis) Role-based access control (RBAC) and data lakes do not seem to go together very well. RBAC controls who can access specific information. Data lakes allow all users to see all information. Is encryption the only way to bridge these two worlds? In this computer science-focused MSc research project and/or master’s thesis, you will investigate existing approaches to managing and enforcing data access in data lakes and explore how encryption can play a central role. https://dasya.itu.dk/for-students/proposals/archive/pito-ml-scheduling/ Thu, 05 Jun 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-ml-scheduling/ Workload collocation has been shown as an effective method to reduce the hardware requirements for certain deep learning (DL) training tasks. On the other hand, there hasn’t been many robust open-source implementations of schedulers that incorporate workload collocation on GPUs for DL. BLOX is a framework that aims at standardizing the way we implement deep learning schedulers. In this project, our goal is to investigate ways to integrate a collocation-aware scheduler in this framework. https://dasya.itu.dk/for-students/proposals/archive/pito-ml-checkpointing/ Thu, 05 Jun 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-ml-checkpointing/ Deep learning training can run for hours or days, causing long queue times and poor quality of service (QoS) on shared clusters. Since schedulers can’t accurately predict training durations, they often wait for jobs to finish or time out, worsening the issue. Frameworks like TensorFlow and PyTorch support model checkpointing, but frequent checkpoints can slow training, while infrequent ones reduce effectiveness. Automating checkpointing could improve scheduler decisions and fairness. This project aims to (1) benchmark checkpointing overhead in TensorFlow and PyTorch, (2) identify optimal checkpointing frequencies, and (3) develop an automatic checkpointing module. https://dasya.itu.dk/for-students/proposals/archive/pito-ml-dataset/ Thu, 05 Jun 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-ml-dataset/ This project focuses on extending an existing dataset for predicting GPU memory requirements during deep learning training by incorporating transformer-based models such as BERT, GPT, and their variants. The student will study the architecture of these models and develop training scripts to run them under controlled conditions. During training, key GPU metrics—including memory usage, utilization, and runtime characteristics—will be recorded. These observations will be used to build a richer dataset that links model specifications with their GPU resource needs. https://dasya.itu.dk/for-students/proposals/archive/pito-ml-gpu_utilization/ Thu, 05 Jun 2025 00:00:00 +0000 https://dasya.itu.dk/for-students/proposals/archive/pito-ml-gpu_utilization/ GPU offers massive computational power and parallelism through its Streaming Multiprocessors (SMs). Efficient GPU utilization is critical for maximizing performance and optimizing compute resource usage, which is measured using various metrics such as SMACT (SM Activity) and SMOCC (SM Occupancy), and DRAMA (DRAM Active). These metrics provide insight into how effectively the GPU’s SMs and memory are used during deep learning tasks. Currently, monitoring tools provide such metrics during runtime. This project aims to explore the concept of predicting GPU utilization metrics using deep learning models, trained on a large dataset of models executed on NVIDIA A100 40GB GPUs.