Thesis

Distributed Stream Processing Systems (DSPS) allow real-time processing of data to fetch meaningful information. Big giants like Twitter, Facebook, and Alibaba rely on DSP for real-time data analytics. For example, click analytics, credit card fraud detection, etc. In DSPS, queries are usually long-running. They typically deal with a very high workload of millions or even billions of events per second. Under such scenarios, parallelism plays an important role in providing scalability for DSPS where either the input stream is partitioned into multiple processes or sub-streams or duplicated to be processed by operators and their instances. However, existing DSPS lacks the benchmarking of the parallel processing capability of operators under various workloads, event rates, and different queries.

Research Goals:

• Literature study of benchmarking parallel stream processing.
• Implementing Window and Pane-based parallelization strategy.
• Evaluating the performance for various benchmarks
• Displaying the performance evaluation result on the front end.
• Possible DSPS to be used: Apache Flink

Ein resilientes Kommunikationssystem soll immer verfügbar und zuverlässig sein, wie erwaretet funktionieren und sich schnell und selbstständig erholen, wenn es zu Störungen oder Angriffen kommt.

In meiner Forschung geht es darum, die Resilienz von aktuellen und zukünftigen mobilen Zugangsnetzen (5G und Beyond-5G) zu verbessern.

Der Kern eines 5G-Netzes besteht aus verschiedenen virtualisierten Netzfunktionen. Dadurch ergeben sich verschiedene Möglichkeiten diese Netzwerkfunktionen zu nutzen. Beispielsweise können sie verteilt auf mehreren VMs platziert oder in K8s-Pods bereitgestellt werden.

Eine ausfallsichere Lösung sollte dann in der Lage sein, den Absturz und den anschließenden Neustart einzelner Netzwerkfunktion zu bewältigen, ohne dass das Gesamsystem übermäßig beeinflusst wird. Desweiteren sollte das Kernnetz in der Lage sein, einen effizienten Lastausgleich zu ermöglichen und sich an Änderungen anzupassen, indem die Anzahl der Netzwerkfunktionen erhöht oder verringert wird.

Ich biete Bachelor- und Masterarbeiten an, die sich mit dem 5G-Kernsystem befassen und Aspekte im Zusammenhang mit dessen Resilienz untersuchen.

Motivation

 In Distributed Stream Processing Systems (DSPS), queries are usually long-running. They typically deal with a very high workload of million or even billion of events per second. Under such scenarios, parallelism plays an important role in providing scalability for DSPS. One of the core
decisions here is to determine the right parallelization degree to process a high load of events and queries while ensuring high throughput and low latency requirements. In this context, parallelization can be controlled by defining the degree of parallelism for each operator in the operator graph.
The main questions we want to investigate here are - 

•  What is the right parallelism degree for query processing? (what is the performance in terms of end-to-end latency for parallel stream processing? At what extend the model can be used - seen and unseen behaviour? What are the limitations of models? At which can be extended?)
• How to parallelize the processing of DSPs operators? 

 Prerequisites

 There are no hard guidelines for this topic. However, it is preferable if you have:

• Good programming skills in Java and Python
• Good understanding of the concepts of machine learning
• Understanding of the concept of communication networks, big data processing engines, e.g., Apache Flink, Kafka. 

 Reference Literature

• Three steps is all you need: fast, accurate, automatic scaling decisions for distributed streaming dataflows.
• A Comprehensive Survey on Parallelization and Elasticity in Stream Processing
• Resource Management and Scheduling in Distributed Stream Processing Systems: A Taxonomy, Review, and Future Directions
• One Model to Rule them All: Towards Zero-Shot Learning for Databases Benjamin
• Apache Flink documentation: nightlies.apache.org/flink/flink-docs-master/docs/learn-flink/overview/ [https://nightlies.apache.org/flink/flink-docsmaster/docs/learn-flink/overview/] 

Contact:

I am looking for motivated students interested in working on cutting-edge technologies and IoT applications area. If you are interested in writing a Bachelor's or Master's Thesis or gaining experience in MAKI project as HiWi, please feel free to contact me. You can email me the following information:

• Your CV or small text about your courses and prior experiences in this area.
• The short note on your motivation about this topic and,
• Transcript or TUCaN grade list.

Overview

The Interplanetary Filesystem (IPFS) is a P2P data storage and retrieval network. Structurally, it exhibits features of both structured and unstructured networks. A Kademlia-based DHT is used to store lists of providing peers for data items. Requests for data are first locally flooded in an unstructured fashion, and only on failure located through the DHT and scoped more precisely.

We have multiple years of recorded requests from the IPFS network, including information about the clients and geolocation of the request origins. This is a large dataset that needs to be analyzed. Furthermore, data collection is an ongoing process and can be augmented, if necessary.

Goal

Derive useful information from the dataset.

Tasks

1. Understand the dataset and how we capture it. Understand its limitations.
2. Come up with interesting queries about the dataset.
3. Engineer solutions that incrementally (i.e., batch-processing new data) derive answers to these queries from the dataset.
4. Implement said solutions to automatically(!) derive insights into the dataset and visualize them.

Requirements

• Proficiency with Linux and familiarity with the command line. Data analysis will almost exclusively run on remote servers due to the volume of the data.
• Motivation about the topic. Don't pick this topic if all you need is a thesis project.
• Knowledge of common formats such as JSON and CSV
• Knowledge of Python (and R) for data processing and visualization. Final visualizations etc. should be done using R, but it's possible to learn enough R in half a year to make this work.

Literature

• https://arxiv.org/abs/2104.09202
• https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9142764

Motivation

In der Forschungsgruppe Adaptive Communication Systems (ACS) beim Lehrstuhl Multimedia Communications (KOM) beschäftigen wir uns u.a. mit der Frage, wie sich Kommunikationssysteme an aktuelle Rahmenbedingungen im Netzwerk anpassen lassen. Die Ziele dieser Adaption reichen von der Verbesserung der Dienstgüte (Quality of Service) in Katastrophenszenarien bis hin zur Verbesserung der Nutzererfahrung (Quality of Experience) im alltäglichen Leben. Im Kern dieser Adaption lässt sich Machine Learning (ML) anwenden, beispielsweise um das Verhalten von Netzwerkknoten ohne manuelle Vorgaben dynamisch an neue Rahmenbedungen anpassen zu können.

Die Kombination von Reinforcement Learning (RL) und Deep Learning (DL) ermöglicht in letzter Zeit große Erfolge, insbesondere auch im Netzwerkbereich. Motiviert durch die Adaption von Kommunikationssystemen betrachten wir im Rahmen dieser Ausschreibung das Lernen innerhalb von Systemen mit mehreren Agenten (Multi-Agent Reinforcement Learning, MARL) unter Unsicherheit und kooperativen Zielen. In einer Thesis werden Teilaspekte dieser Problemstellung behandelt.

Mögliche Schwerpunkte

Das spezifische Thema der Thesis können wir gemeinsam und in Abhängigkeit von deinen Interessen definieren. Schreib mir dazu einfach eine E-Mail, in der du dich kurz vorstellst. Aktuell sind Themen mit den folgenden Schwerpunkten möglich:

Wenn dich davon etwas anspricht, können wir konkretere Themenvorschläge zu den Schwerpunkten besprechen. Alternativ kannst du auch eine eigene Problemstellung vorschlagen, die Überschneidungen mit diesen Bereichen aufweist. Beispielsweise eine für dich interessante netzwerktechnische Problemstellung oder ein (Video)Spiel, für welches du eine AI entwickeln möchtest. Kooperation und Kommunikation sind oft zentrale Aspekte von verteilten Systemen.

Anforderungen

Literatur zum Einstieg

Allgemeine Literatur zu (Deep) Reinforcement Learning in Multiagentensystemen und Netzwerken:

[1] P. Hernandez-Leal, B. Kartal and M. E. Taylor, "A survey and critique of multiagent deep reinforcement learning," in Autonomous Agents and Multi-Agent Systems, vol. 33, no. 6, pp. 750-797, 2019.
[2] N. C. Luong et al., “Applications of Deep Reinforcement Learning in Communications and Networking: A Survey,” IEEE Communications Surveys Tutorials, vol. 21, no. 4, Art. no. 4, 2019.

Zum Einstieg in RL ist Q-Learning im RL-Standardwerk [3] und Deep Q-Learning [4] zu empfehlen

[3] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. MIT press, 2018.
[4] V. Mnih, et al, "Human-level control through deep reinforcement learning," in nature, vol. 518, no. 7540, pp. 529-533, 2015.

Zum Einstieg in Kommunikation ist TarMAC [5] empfehlenswert, eine Übersicht bietet das Survey von Zhu et al. [6]

[5] A. Das et al., “TarMAC: Targeted Multi-Agent Communication,” in Proceedings of the 36th International Conference on Machine Learning, vol. 97, pp. 1538–1546. [Online]. Available: http://proceedings.mlr.press/v97/das19a.html
[6] C. Zhu, M. Dastani, and S. Wang, “A Survey of Multi-Agent Reinforcement Learning with Communication.” arXiv, 2022. doi: 10.48550/ARXIV.2203.08975.

Motivation:

Internet of Things describes a new world of heterogeneous objects such as sensors, smartphones, actuators etc, which intelligently interacts and communicate with each other and offer a better quality of life. With the increase of these objects, the data volume is growing as well. Processing this volume of data can be helpful to various application domains to reveal complex patterns, gain fast insights or react to observed situations, e.g., fraud detection, smart factories, telecommunication, etc. However, it becomes extremely challenging to continuously process this high volume of data in a timely manner and meet these applications' demands. One of the possible ways can be to use heterogeneous resources' capability such as GPUs, FPGAs, etc. to accelerate the big data analytics instead of relying on homogeneous resources, e.g., CPU.

Task:

• Analyze existing related literature.
• Formalizing the problem and developing the solution and
• Setting up development environment and implementing the proof of concept or prototype.
• Evaluation of the proposed solution in comparison to the existing approaches. .

Prerequisites:

There are no hard guidelines for this topic. However, it is preferable if you have:

• Good programming skills in Java and/or Scala.
• Understanding of the concept of .
• communication networks and .
• big data processing engines, e.g., Apache Flink, Kafka.

Contact:

I am looking for motivated students interested in working on cutting-edge technologies and IoT applications area. If you are interested in writing a Bachelor's or Master's Thesis or gaining experience in MAKI project as HiWi, please feel free to contact me. You can email me the following information:

• Your CV or small text about your courses and prior experiences in this area.
• The short note on your motivation about this topic and,
• Transcript or TUCaN grade list.

To deal with rapidly changing workloads,  Distributed Stream Processing Systems rely on parallelism to use multiple instances of operators to process a high amount of data. In recent times, several heuristic-based and learned-based approaches have been proposed to tune parallelism, but they are limited to specific workloads and can't work for unseen workloads which could result in degraded performance and inefficient resource utilization. In this thesis, we look at how can end-to-end learning model be  used for parallelism prediction in distributed systems.  The proposed model should aim to accurately predict optimal parallelism levels under a variety of scenarios, thus improving data processing efficiency and resource utilization in distributed systems. 

Graph Neural Networks (GNNs) have shown potential in various domains, including databases and distributed stream processing systems. However, the effectiveness and accuracy of GNN models, particularly in zero-shot learning scenarios, heavily depend on selecting appropriate features. This thesis aims to develop and evaluate feature selection strategies and an optimized method to select from feature selection strategies based on constraints that enhance the performance of zero-shot-based GNN models in the context of parallel stream processing.

Key Objectives

• Identify and analyze the key features that impact the performance of GNN models in parallel stream processing tasks.
• Develop a systematic feature selection strategy that optimizes the zero-shot learning capabilities of GNN models in distributed stream processing environments.
• Evaluate the effectiveness of the proposed feature selection strategy in improving the accuracy, efficiency, and scalability of zero-shot GNN models.

In recent years, there has been an increasing interest in adopting federated learning to train deep learning models. 
This heightened interest is primarily driven by the agility, data privacy, scalability, and efficiency that federated learning has demonstrated.
Federated learning (FL) has gained significant traction, particularly in developing and updating object detection models.
These models demand continuous refinement, a task that FL can efficiently accomplish. Nonetheless, a critical challenge for these models revolves around the collection and preparation of training data.
This issue is particularly pronounced in vehicular networks, where each vehicle or node must gather images and prepare them for their respective local models.
However, a pivotal prerequisite is labeling these images before they can be utilized in training supervised models.
This thesis presents several methodologies for image labeling, offering an in-depth exploration of each approach, and showing their respective advantages and drawbacks (e.g., in the model performance, required computational power, and network load).
Furthermore, it conducts a comprehensive assessment of the performance and accuracy of these approaches, utilizing a variety of metrics for evaluation.

Task

• Explore federated learning applications, with a particular focus on their utilization within the domain of vehicular applications. 
• Extend the implementation of the pre-existing framework for training federated learning-based object detection models. This extension should encompass the integration of diverse data labeling methodologies.
• Conduct a thorough and exhaustive evaluation of the performance of these aforementioned approaches. This assessment should encompass the utilization of a diverse array of metrics for evaluation purposes, ensuring a comprehensive understanding of their respective strengths and weaknesses.
• Report the findings

Prerequisites

It is required to have:
•    Motivated and individual working style
•    Very good programming skills in Python.
•    Very good machine learning / Deep learning knowledge

Contact

I am looking for motivated students interested in working on cutting-edge technologies and vehicular applications area. If you are interested in writing a Bachelor's or Master's Thesis, please feel free to contact me at Ahmad Khalil (ahmad.khalil(at)kom.tu-darmstadt.de). You can email me the following information:
•    Your CV or small text about your courses and prior experiences in this area.

Das am Fachgebiet KOM entwickelte OpenSource Framework P4STA (https://github.com/ralfkundel/P4STA) erlaubt hochgenaue Latenzmessungen mit Nanosekundengenauigkeit bei Bandbreiten von bis zu 100 Gbit/s. Um dies zu erreichen werden softwarebasierte Standardlastgeneratoren mit einem programmierbaren Netzwerkswitch kombiniert.

Nachteil hierbei ist, dass die Softwarelastgeneratoren zusätzliche Server benötigen und nur eingeschränkt flexibel sind. Spezielle Netzwerkprotokolle wie beispielsweise PPPoE oder GTP, was in 5G Netzwerken zum Einsatz kommt, können nur schwierig erzeugt werden.

Ziel dieser Arbeit ist es, dass P4STA framework um integrierte Paketgenerierung innerhalb des programmierbaren P4-Switches zu erweitern.
Der Vorteil dieser hardwarebasierten Lastgenerierung liegt zum einen in den hohen Datenraten; pro Port können so bis zu 100 Gbit/s erzeugt werden.
Außerdem können beliebige Protokollstacks erzeugt werden.

Voraussetzungen:

- Gutes Verständnis von Netzwerken und Protokollen
- Vorlesung Software Defined Networking (oder vergleichbar)
- idealerweise Grundkenntnisse in der P4 oder FPGA Entwicklung
- Erfahrung in der Softwareentwicklung, idealerweise z.B. Python/Django/HTML/Javascript

Forschungsfragen, die im Rahmen dieser Arbeit beantwortet werden:

• wie flexibel kann mit hardwarebasierten Lastgeneratoren Last erzeugt werden?
• wird durch eine integrierte Lasterzeugung die Genauigkeit beeinflusst?
• ... weitere Fragestellungen, die im Rahmen der Bearbeitung herausgearbeitet werden ...

Problem Description

We maintain a crawler for libp2p-based Kademlia DHTs, which we use to crawl IPFS.
We are able to enumerate all DHT servers, i.e., all nodes participating as server nodes in the DHT.
We can then visualize data about these nodes, such as agent versions, churn, geolocation, etc.

The crawler, however, is generic over all DHTs implemented using go-libp2p-kad-dht.
There are many projects using this implementation, which we could potentially crawl.
They differ in a) their bootstrap nodes, and b) their protocol identifiers.

Main Objective

Working on well-connected Linux servers, set up multiple instances of the crawler, enumerating peers of as many networks as possible.
Set up automated data wrangling and visualization pipelines, displaying the results on a statically-generated website.

Optional Bonus Goals

• Extend the visualization and data science scripts.
• Realize the setup in a containerized environment, in a way that is easily extensible with new networks.
• Save the results of the crawls into a PostgreSQL database for compact storage and easier queries.
• ???

Prerequisites

You must

• Have intrinsic motivation to look at P2P networks.
• Be able to fluently work on remote Linux machines via SSH.
• Understand Go code.
• Be able to admin Linux machines and document the setup.

It would probably be helpful if

• You understood how DHT crawlers work in general.
• You knew Chinese, since many of the projects have Chinese source code comments or documentation.
• You knew R, since the evaluation scripts are in R.
• You were well-versed in SQL and design of database schemas.

Bachelor/Master?

If you pursue this as a Bachelor's thesis, the minimum goal and any number of optional goals apply.

If you pursue this as a Master's thesis, you'll have to significantly extend the data processing and visualization frameworks to derive new interesting data from the datasets. Addiitonally, you'll of course have to implement the setup in a clean way, well-documented, with more of the optional goals realized. I'd also expect you to come up with your own optional goals.

Motivation

Object detection is a crucial task in the field of autonomous vehicles. To train the object detection models, data such as RGB data is collected either offline or online during the operation of the cars, and this data is used to train the models. However, in supervised learning approaches, which have superior performance, the data needs to be annotated.

Active learning is an approach that aims to reduce the amount of annotated data required. It achieves this by selecting the most informative information from the unannotated data pool. Unfortunately, most active learning approaches for object detection assume that the data is already gathered and located in a central entity (offline). These approaches overlook the consumption of networking resources that occur when transmitting the data in the online data gethering.

In this thesis, the student will have the chance to enable online data collection in a multi-agent system (e.g., vehicular system), while considering the available network resources. The objective is to develop an active learning approach that optimizes the utilization of the network resources while maintaining high object detection performance. This would involve selecting the most informative data for annotation and training, taking into account the limitations of the available network resources.

The goal is to enhance the efficiency and effectiveness of object detection in autonomous vehicles by developing an active learning approach that leverages the available network resources without compromising the performance of the models.

Task

This thesis is open for master/bachelor students, and the main tasks are as follows:
•    Explore relevant existing related literature.
•    Formalizing the problem and modeling the solution and
•    Setting up a development environment.
•    Evaluation of the proposed solution in comparison to the existing state-of-the-art approaches.

Prerequisites

It is required to have:
•    Motivated and individual working style
•    Very good programming skills in Python
•    Very good machine learning / Deep learning knowledge

Related Literature

[1] Yu, Weiping, et al. "Consistency-based active learning for object detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.

[2] Choi, Jiwoong, et al. "Active learning for deep object detection via probabilistic modeling." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.

[3] Feng, Di, et al. "A review and comparative study on probabilistic object detection in autonomous driving." IEEE Transactions on Intelligent Transportation Systems 23.8 (2021): 9961-9980.

Motivation 

Object detection holds paramount importance in vehicular applications, serving as a cornerstone for road safety, efficient traffic management, and the evolution of autonomous driving systems. However, while training the object detection models in vehicular applications, the intricacies of data privacy, constrained communication bandwidth, and computational limitations present significant challenges. This thesis aims to leverage the potential of adaptive model compression techniques to augment Federated Learning-based object detection model training in vehicular applications. The central premise is to develop compression strategies (such as pruning, and quantization) that dynamically respond to contextual information, such as weather conditions, traffic density, and time of day, to optimize model training efficiency and accuracy. This integration holds the potential to create more resource-efficient model training that can adapt to changing environments and perform optimally under varying conditions.

Task

This thesis is open for master's students, and the main tasks are as follows:
•    Explore relevant existing related literature.
•    Formalizing the problem and modeling the solution and
•    Setting up a development environment and implementing the proof of concept or prototype.
•    Evaluation of the proposed solution in comparison to the existing state-of-the-art approaches.

Prerequisites

It is required to have:
•    Motivated and individual working style
•    Very good programming skills in Python.
•    Very good machine learning / Deep learning knowledge

Contact

I am looking for motivated students interested in working on cutting-edge technologies and vehicular applications area. If you are interested in writing a Bachelor's or Master's Thesis, please feel free to contact me at Ahmad Khalil (ahmad.khalil(at)kom.tu-darmstadt.de). You can email me the following information:
•    Your CV or small text about your courses and prior experiences in this area.

• Design und Implementierung geeigneter Mechanismen für den Betrieb redundanter UPFs im 5G Core
• Evaluation der Auswirkungen der Migration von User Sessions zu replizierten UPFs
• Evaluation verschiedener Kommunikationsstrategien zwischen duplizierten UPFs und
  anderen Netzwerkfunktionen

Motivation 

As the development of autonomous vehicles continues to progress, the need for improving detection models becomes increasingly important. However, sharing the data gathered from vehicles with a central server can raise privacy concerns and consume a significant amount of network capacity. Federated learning provides a solution to this challenge by allowing for the training of machine learning models on decentralized data, without the need to transfer raw data to a central server. However, federated learning inherits convergence issues, especially when the data is non-iid (non-independent and identically distributed). This makes studying how the data distribution influences the performance of perception models trained in a federated learning setup in the vehicular collective perception field crucial. By analyzing state-of-the-art approaches and developing novel methods that take into account the data heterogeneity, the student will help improving the accuracy and reliability of perception models. Moreover, training models locally in the vehicle poses significant challenges that require careful consideration, such as dealing with the labeling issue of data collected in real-time. The student will have the opportunity to address these challenges and contribute to the future of autonomous vehicles.

Task

This thesis is open for master's students, and the main tasks are as follows:
•    Explore relevant existing related literature.
•    Formalizing the problem and modeling the solution and
•    Setting up a development environment and implementing the proof of concept or prototype.
•    Evaluation of the proposed solution in comparison to the existing state-of-the-art approaches.

Prerequisites

It is required to have:
•    Motivated and individual working style
•    Very good programming skills in Python.
•    Very good machine learning / Deep learning knowledge

Contact

I am looking for motivated students interested in working on cutting-edge technologies and vehicular applications area. If you are interested in writing a Bachelor's or Master's Thesis or gaining experience in MAKI project as HiWi, please feel free to contact me at Ahmad Khalil (ahmad.khalil(at)kom.tu-darmstadt.de). You can email me the following information:
•    Your CV or small text about your courses and prior experiences in this area.

Motivation

Time-Sensitive Networking (TSN) is an evolving set of standards for reliable real-time ethernet communication.

It's possible to deploy current or future TSN mechanisms on hardware switching devices for reliable Ethernet communication. The Co-existence of multiple TSN mechanisms is however still challenging problem, as complex interplay between different approaches can occur.

Tasks

1. Review related work of time-sensitive networking mechanisms and simultaneous deployments
2. Design approaches to enable simultaneous deployments
3. Implement the approach on test-devices
4. Evaluate your deployment
5. Write a thesis

Requirements

1. Basic computer networking knowledge
2. Interest in real-time networking