Overview:
LLM-based intelligent agents face significant deployment challenges, particularly related to resource management. Allowing unrestricted access to LLM or tool resources can lead to inefficient or even potentially harmful resource allocation and utilization for agents. Furthermore, the absence of proper scheduling and resource management mechanisms in current agent designs hinders concurrent processing and limits overall system efficiency.
Simulation provides a controlled and reproducible environment to study these challenges in LLM-based agent systems across diverse workflow mixes and hardware configurations, without the cost and variability of real deployments. By designing and implementing models of agentic workflow behavior in an existing systems simulation framework, concurrent LLM calls, tool invocations, and background tasks can be studied systematically, allowing for testing novel abstractions for resource-allocation or scheduling strategies.
The central question of this thesis is how the behavior of agentic workflows can be characterized and modeled in a form that allows existing system simulators to represent their execution. The core idea is to build on established workload abstractions while extending their execution models to capture agent-specific behavior, dynamic decision-making, and resource requirements that are not currently represented.
Research Questions:
Which execution and resource-usage characteristics distinguish agentic workflows from conventional workloads, and which of these need to be represented in system simulators?
How can the dynamic behavior of agentic workflows, such as task generation, branching, tool invocation, and changing resource demands, be characterized and modeled from execution traces?
How can these models be integrated into existing simulation frameworks such as WRENCH, SimGrid, or CloudSim while preserving their established workload and resource abstractions?
Requirements:
Solid knowledge of operating systems concepts and advanced programming skills in C++ and Python or similar as well as foundational knowledge on AI systems are required. The willingness to learn about unfamiliar systems, tools and methods is expected. Your learning process will be highly rewarding.
Start: Immediately
Contact: Niklas Fomin (fomin ∂ tu-berlin.de)
Note: If applicable, we expect the results to be used for further research and publication in cooperation with the student (if wanted).
Resources:
Github
https://github.com/wrench-project/wrench
Papers