Scalable Parallel Computing Laboratory

Mentors: Student: luca-patrignani

The DaCe framework uses the Stateful DataFlow multiGraph (SDFG) data-centric intermediate representation. This graph maps to a variety of high-performance architectues and performance engineers can manipulate this graph for achieving state-of-the-art representations. When dealing with large projects with hundreds of thousands lines of codes the state graph can become extremely unreadable due multiple nested cycles, function’s calls, and conditional branching: it becomes nearly impossible to understand smaller parts of the graph and it also makes relating it to the original source code very hard. To mitigate this problem the concept of ControlFlowRegion was introduced. A ControlFlowRegion is a region of the SDFG which represents a control flow construct just like a traditional programming language. An example can be the LoopRegion, developed on a separate feature branch (loop_architecture_pt_3), which represents the classic C-like for loop for ( init-clause ; cond-expression ; iteration-expression ). The project goal is to - Implement all the other control flow constructs like if … elif … else and match - Implement regions for function calls and with statements - Implement user defined regions - Add support for all these new regions in the DaCe VSCode extension

Mentors: Student: Prajin

The project aims to enhance the Serverless Benchmark Suite (SeBS) by implementing multiple invocation mechanisms, complex workflows, and container deployment support offering a comprehensive evaluation of serverless architectures across major cloud platforms. This project will expand SeBS with queue and storage-based invocations, integrate standalone benchmarks across a variety of tasks, and implement real-world complex workflows such as ETL processes, machine learning inference, and training. And, extend SeBS to support container-based deployment along with zipped deployment. These enhancements are designed to provide in-depth analysis and insights into the performance and scalability of serverless platforms.

Mentors: Student: Syed Mujtaba

FaaSKeeper is a ZooKeeper-like serverless coordination service. It implements standard ZooKeeper functionalities and includes a new client for Python applications. ZooKeeper has been used by many Apache projects. The goal of this project proposal is to identify an Apache project that can use FaaSKeeper instead of Zookeeper, define the Zookeeper APIs that will have to be implemented in FaasKeeper (if any) and create a new Java client library that offloads ZooKeeper calls to the REST cloud API of FaaSKeeper instead of using the ZooKeeper protocol.

Mentors: Student: AbhishekKumar

The serverless computing landscape is constantly evolving with new types of applications and use-cases. To ensure the SeBS benchmarking suite remains relevant and comprehensive, this project aims to extend it by introducing new function candidates, invocation types, and serverless workflows. The project's objective is to develop and integrate new benchmarks that can be executed successfully on already supported serverless platforms.

Mentors: Student: Fu-Chiang

In this project, my job is to extend the existing Dace Distributed framework to support MPI4PY syntax. This will mainly involve extending the existing Dace parser to support MPI4PY syntax, as well as implementing MPI I/O and Comm World functionality. The entire implementation will be validated through a series of unit tests and mentor evaluations.