Prof. Dr. Frank Bellosa

Abstract
Distributed systems primitives, such as consensus, play an important role in current software deployments. Especially in datacenters, where applications are split between different servers, mechanisms for coordination are required. The libraries and applications which provide such functionality currently execute alongside other applications and the operating system on the host. As a consequence of task switching and system calls, they are subject to unpredictable spikes in latency, which adversely affects performance. Recent integrations of FPGAs into PCIe networking cards present an opportunity in this respect. Due to increased proximity to the network transceivers and operational model of FPGAs, they experience lower and more stable latencies. To leverage this advantage, distributed systems primitives must be located on the FPGA itself.
This thesis analyses the communication latency and bandwidth characteristics provided by these accelerators in a datacenter setting. It then develops and implements an abstraction for synchronous algorithms to be implemented against. To show the viability of the framework and determine its performance characteristics, this thesis further implements a consensus primitive based on existing algorithms to leverage the developed  bstraction. In order to evaluate the performance of the algorithm, as well as the framework components, a number of micro-benchmarks as well as full-system experiments are run. Results show that the framework is able to saturate 100G links while introducing minimal overhead, both in latency as well as required FPGA resources. This enables a new class of synchronous algorithms to be implemented with both reduced effort and a lower barrier to entry.

BibTex:

@masterthesis{Deterministic PU (detPU),
  author = {Philipp Ludwig Waldemar Berdesinski},
  title = {Deterministic PU (detPU): FPGA-based fast, flexible coordination for datacenters},
  type = {Master Thesis},
  year = 2025,
  month = jan# "02",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
  }