Current Theses

Title	Student	Supervisor	Description
Dynamic Scheduling of Serverless Functions	Christian Chisté	Fedor Smirnov	details
A learning-based simulation framework for volatile cloud resources	Johannes Spies, Simon Triendl	Sashko Ristov	details
Decoupled Automatic Deployment of AFCL Workflows	Caroline Haller, Christoph Abenthung	Sashko Ristov	details
Function Choreography Scheduling Framework for Multiple FaaS Systems	Tobias Pockstaller	Sashko Ristov	details
High-level Distributed GPU Programming using Celerity	Ralf Pernecker, Linus Wald	Philipp Gschwandtner	details
G2GA: Portable execution of workflows in Google Cloud Functions across multiple FaaS platforms	Anna Kapeller, Felix Petschko	Sashko Ristov	details
Running workflow applications across multiple cloud providers	Marina Aichinger	Sashko Ristov	details
Interoperable and secure execution of function choreographies in a heterogenous edge-cloud continuum	Lucas Markovic	Sashko Ristov	details
Memory Management and Page Replacement Algorithms in the x86_64 Architecture	Stefan Kuhnert	Thomas Fahringer	details
Establishing Virtual Networks in Amazon Web Services	Hüseyin Gündogan	Sashko Ristov	details
Code Generation and Specialization for Unstructured Mesh Stencils	Claudia Laube	Peter Thoman	details
Fluid particle simulation using SYCL	Markus Wippler	Peter Thoman	details
Streaming of Voxel-based Geometry on GPUs	Philipp Mildenberger	Peter Thoman	details

If a bachelor student wants to set his/her initial/final presentation he/she (or the supervisor) MUST contact Sashko Ristov to schedule the presentation!

Details for the theses

Title	Dynamic Scheduling of Serverless Functions
Student	Christian Chisté
Language	English
Supervisors	Fedor Smirnov
Description	The execution of applications which are distributed over a large number of cloud resources is subject to a high degree of dynamism w.r.t., e.g. dynamic resource failures or a dynamic variation of the task execution time. Since this dynamism is very difficult or even impossible to predict, generating static schedules which, prior to the execution of the workflow, already fix the task-to-resource mapping, is difficult and may lead to situations where the workflow execution must be aborted and restarted, e.g., due to a single resource failure. For these situations, approaches capable of a dynamic rescheduling of individual tasks/parts of the workflow are a promising alternative to purely static approaches. In this Bachelor’s thesis, an existing engine shall be extended by the capability for dynamic communication with a scheduler. The augmented enactment engine will then be used for experiments in which static and dynamic scheduling approaches are compared with each other.
Tasks	Extending the existing engine by adding the ability to communicate dynamically with the scheduler. Design and implementation of an experimental setup to compare static and dynamic scheduling approaches. Evaluation of the experimental results and writing the thesis summarizing the work.
Theoretical skills	Distributed Systems
Practical skills	Java
Additional information

Title	A learning-based simulation framework for volatile cloud resources
Student	Johannes Spies, Simon Triendl
Language	English
Supervisors	Sashko Ristov
Description	Cloud simulation frameworks can simulate the behavior of complex applications when being executed on cloud resources. However, the simulators need to be configured with many parameters which are not known by the users. Even a small inaccuracy in the simulation may generate a huge discrepancy between the simulated and real behavior. In this theses, the students will extend an existing simulation framework to improve simulation accuracy based on insights gained by real execution and performance measurements and use the same application and resource setup for the simulation.
Tasks	Extend the VolatileSim simulator with an execution module for DAX execution on AWS EC2 Measure performance of EC2 instances for various workflow application tasks Configure the simulation framework with knowledge gained by running workflows on EC2 Collect and analyse resulting data
Theoretical skills	Distributed Systems, workflow applications, Scheduling, DAX
Practical skills	AWS, Java.
Additional information	The student will work on the VolatileSim simulation framework, which was developed as a part of the master thesis “VolatileSim: a simulation framework for cloud volatile resources”, written by Christoph Schöpf MSc, supervised by Dr. Sashko Ristov.

Title	Decoupled Automatic Deployment of AFCL Workflows
Student	Caroline Haller, Christoph Abenthung
Language	English
Supervisors	Sashko Ristov
Description	This bachelor thesis will develop a system that will deploy serverless functions of an AFCL workflow including required cloud services across multiple FaaS regions.
Tasks	User Interface for function types and implementation meta-data Automatic deployment of a function implementation with dependent cloud services Decoupled deployment scripts for a given CFCL file on multiple cloud regions Evaluation of the automatic deployment of various function implementations of a single function type
Theoretical skills	Distributed Systems, workflow applications, serverless computing
Practical skills	Java, SDK / API of public cloud providers.
Additional information	Both the AFCL API (in Java) to build the function choreography, a parser and the enactment engine to run function choreographies (serverless workflows) are developed by the DPS group and will be provided to students.

Title	Function Choreography Scheduling Framework for Multiple FaaS Systems
Student	Tobias Pockstaller
Language	English
Supervisors	Sashko Ristov
Description	This bachelor thesis will develop a framework that will adapt and schedule a function choreography for portable execution across multiple resources, including all regions of multiple public FaaS providers (e.g. AWS and IBM), as well as the on-premise serverless platforms (e.g. OpenWhisk).
Tasks	Develop AFCL – Opt4J (input) and OPT4J – CFCL (output) interfaces Integrate an AFCL-AFCL adaptor for parallel loop unrolling. Develop a knowledge base Develop and deploy a real life function choreography on at least two FaaS systems (e.g. AWS and IBM) Evaluate the quality of scheduled solutions by running the function choreography across multiple FaaS systems (e.g. AWS and IBM).
Theoretical skills	Distributed Systems, workflow applications, serverless computing
Practical skills	Java, SDK / API of public cloud providers.
Additional information	Both the AFCL API (in Java) to build the function choreography, a parser and the enactment engine to run function choreographies are developed by the DPS group and will be provided to the student.

Title	High-level Distributed GPU Programming using Celerity
Student	Ralf Pernecker, Linus Wald
Language	English
Supervisors	Philipp Gschwandtner
Description	GPUs are an inreasing source of computational power in today’s high performance computing community. Already more than half of the current top 50 supercomputers worldwide are equipped with GPUs. This bachelor thesis focuses on evaluating the Celerity programming framework, a combined API and task-based runtime system for high-level distributed GPU programming, by implementing several benchmarks and real-world applications and evaluating their performance on a medium-sized GPU cluster.
Tasks	implement several benchmarks and real-world applications using the Celerity programming framework benchmark these applications on a GPU-based distributed memory cluster evaluate the results and Celerity’s effectiveness for real-world applications
Theoretical skills	Parallel Programming, High Performance Computing
Practical skills	C++, GPU Programming, working on HPC clusters
Additional information

Title	G2GA: Portable execution of workflows in Google Cloud Functions across multiple FaaS platforms
Student	Anna Kapeller, Felix Petschko
Language	English
Supervisors	Sashko Ristov
Description	Public FaaS providers limit function execution in multiple ways (duration, assigned memory/CPU, input/output size, code size, etc). Moreover, each widely used FaaS provider offer a language to compose a workflow of serverless functions. However, they lock their users into their own platform. To overcome these limits, this bachelor thesis will develop a system for portable execution of workflows developed in one serverless workflow provider (e.g. Google Air Flow) across multiple FaaS platforms (e.g. Google and Microsoft).
Tasks	Develop invokers for portable execution of serverless functions across multiple FaaS providers (e.g. Microsoft Azure and Google Cloud Functions) Develop a converter from Apache Airflow to AFCL to enable a portable execution of workflow applications implemented in Apache Airflow across multiple FaaS providers Deploy two workflow applications (e.g. Montage and 1000genome-workflow) as Serverless workflow applications on multiple FaaS providers (e.g. Azure and Google Cloud Functions) Evaluate the potential benefit of running serverless workflow applications across multiple FaaS providers versus a single FaaS provider (e.g. Google Cloud Functions).
Theoretical skills	Distributed Systems, workflow applications, serverless computing
Practical skills	Java, SDK / API of public cloud providers.
Additional information	Both the AFCL API (in Java) to compose the workflows and a parser are developed by the DPS group and will be provided to the student.

Title	Running workflow applications across multiple cloud providers
Student	Marina Aichinger
Language	English
Supervisors	Sashko Ristov
Description	Running workflow applications across multiple cloud providers offers many benefits as it can overcome limits and constraints of individual cloud providers, various pricing schemes, etc. However, it requires an efficient resource management and integration of various APIs. This project provides an engine that can integrate resources of multiple IaaS providers in the workflow description and run the workflow across these federated resources.
Tasks	extend AFCL to describe resources develop resource provisioner for resource management adapt the CFCL file with provisioned resources develop an invoker for running tasks in IaaS clouds
Theoretical skills	Distributed Systems, Cloud Computing, workflow applications
Practical skills	Java, SDK / API of public IaaS cloud service providers.
Additional information	Both the AFCL language schema (in YAML) and the API (in Java) to compose the workflows are developed by the DPS group and will be provided to the student.

Title	Interoperable and secure execution of function choreographies in a heterogenous edge-cloud continuum
Student	Lucas Markovic
Language	English
Supervisors	Sashko Ristov
Description	Edge devices have limited resources and the system has to offload the computing to a public cloud (serverless FaaS provider). However, according to the security policy, probably the whole communication outside of user premises should be encrypted (e.g. a serverless function that runs at the edge sends data to a function that runs in a public cloud provider). This requires huge development efforts to adapt the application (functions) to the new execution environment. The goal of this bachelor thesis is to abstract the composition of the workflow of serverless functions from the development of the functions themselves. Students will build a system that can run a function choreography (a workflow of serverless functions) in a multi-provider scenario by retaining the user constraints (security constraints and interoperability) without changing functions, but by introducing additional middleware functions. The function choreography is described with the AFCL (Abstract Function Choreography Language) language in YAML, defined and developed by the DPS group.
Tasks	Define security policy per user that will be valid for each function choreography Extend the abstracted view of the function choreography to comply with the security and interoperability issues when a function choreography is executed across different public FaaS providers or outside of the user premises Develop middleware serverless functions (encrypt / decrypt input – output) and interoperability functions (convert input / output types) Update the workflow description by introducing the middleware functions based on the security policy Measure and profile the performance and cost trade-off of the additional middleware functions
Theoretical skills	Distributed Systems, Cloud Computing, Serverless, encryption
Practical skills	Java, SDK / API of public FaaS systems.
Additional information	Both the AFCL language schema and the JAVA API to compose the function choreography are developed by the DPS group and will be provided to students.

Title	Memory Management and Page Replacement Algorithms in the x86_64 Architecture
Number of student	Stefan Kuhnert
Language	English
Supervisor	Thomas Fahringer
Description	Paging and memory allocations changed a lot since the introduction of x86_64 Architecture. The goal of this thesis is to provide an overview of the changes and analyzing metrics that are used for comparison of memory related algorithms.
Tasks	Differences between the 64-bit and the 32-bit Architecture Tool for Page Replacement & Memory Allocation Algorithms Metrics and Comparisons of Memory Related Algorithms Benchmarks and Categorization of memory intensive applications
Theoretical Skills	Operating System Architecture, Algorithms, Memory management
Practical Skills	x86_64 Architecture, C/C++, Linux Kernel
Additonal Information

Title	Establishing Virtual Networks in Amazon Web Services
Student	Hüseyin Gündogan
Language	English
Supervisor	Sashko Ristov
Description	This bachelor thesis aims to develop a web application, that provides a simple GUI for managing cloud resources on Amazon Web Services. Users should be able to create and modify EC2 Instances by defining their amount, type and security rules. The main goal is to design the required resources as simple as possible, in order to create a virtual network of EC2 Instances by submitting a single request.
Tasks	Develop a GUI to design a network of virtual machines Develop an API for managing cloud resources
Theoretical Skills	Familiar with APIs, programming graphical interface
Practical Skills	Java, Web development
Additonal Information

Title	Code Generation and Specialization for Unstructured Mesh Stencils
Student	Claudia Laube
Language	English
Supervisors	Peter Thoman
Description	In this thesis, generating input-data-specific processing code for performing stencil computation on large unstructured meshes is investigated.
Tasks	Familiarization with the topic of unstructured mesh computations Create a code generator which can generate code for at least 3 distinct types of meshes (ie. 2D grids, fully connected, and manually designed from a common source format) Investigate the reasons the performance advantages of specialized code Find methods to reduce compile times while maintaining the performance advantages Evaluate how the relative performance of traditional stencil code and data-specific generated code behaves for different mesh types, node structures, and update steps

Title	Fluid particle simulation using SYCL
Student	Markus Wippler
Language	English
Supervisors	Peter Thoman
Description	This thesis involves the implementation of a GPU-based fluid particle simulation using SYCL, as well as its visualization.
Tasks	Investigate possibilities for SYCL integration with graphics APIs, make an implementation choice and reason about it Investigate data structures proposed in the literature for this task, perform a comparative evaluation, and choose one for your implementation Implement the particle simulation using SYCL, as well as a visualization Evaluate its performance across at least 2 GPU architectures Implement and evaluate the impact of at least 2 optimization strategies

Title	Streaming of Voxel-based Geometry on GPUs
Student	Philipp Mildenberger
Language	English
Supervisors	Peter Thoman
Description	Investigation of the data structures and algorithms required for high-performance rendering of very detailed Voxel scenes on GPU. Includes the development and implementation of a streaming technique.
Tasks	Comparison of Voxels to different Geometry models like a point cloud or triangles, and their use Explanation why streaming of geometry is necessary for extremely detailed scenes Short discussion about different rendering methods & the necessity of the GPU for raycasting Discussion and development of the data structure needed for streaming Development of a streaming technique