Multi-Level Scheduler for Accelerators in Distributed Systems

Context

Embedded systems, for example in applications such as automated driving, require ever-increasing computing power. To cope with these demands, a variety of different accelerators are used. Existing applications are often developed specifically for individual accelerators in this context, and this is especially true for AI accelerators. We are researching concepts to break this lock-in and enable dynamic load balancing.

Tasks

The goal of this thesis is to develop a concept for load balancing between accelerators in a heterogeneous system. The distribution shall be calculated centrally and performed locally. The thesis shall consist of a literature research, the creation of a concept, as well as a prototypical implementation of the concept and an evaluation of it.

Requirements

• You have very good programming skills in Rust, C++ or Python.
• You have a very good command of German or English, both written and spoken.
• Independent and solution-oriented way of working