Name: EDGARD DA CUNHA PONTES
Publication date: 20/12/2024
Examining board:
Name |
Role |
|---|---|
| MAGNOS MARTINELLO | Presidente |
| MOISES RENATO NUNES RIBEIRO | Coorientador |
| RAFAEL SILVA GUIMARÃES | Examinador Externo |
| RODOLFO DA SILVA VILLACA | Examinador Interno |
Summary: This work investigates the efficiency of transporting large volumes of scientific data in
the context of Data-Intensive Science (DIS), focusing on the use of sliced WANs. These
networks offer a promising solution to meet the flexibility and granular control required for
transferring large volumes of scientific data, surpassing the limitations of traditional WANs,
enabling the creation of virtual, isolated and customizable slices for different performance
and security requirements.
The research utilizes the FABRIC Testbed, an experimental platform in the USA, to
simulate realistic DIS network scenarios, exploring two main cases: single-path and multi-
path data transfer in a distributed WAN. The experimental methodology employs iPerf3
to generate and transfer TCP data flows within the network slices, collecting metrics such
as throughput, completion time, and retransmissions to evaluate various configurations
and algorithms. A central focus is the analysis of TCP congestion control algorithms,
comparing BBR and Cubic. The results show that BBR outperforms Cubic in scenarios
with small buffers and short-to-medium RTTs, demonstrating greater efficiency. The study
also investigates the impact of parameters such as buffer size, latency, flow multiplexing,
and bandwidth allocation, highlighting how their optimization maximizes throughput and
reduces transfer time.
In the multi-path context, the research examines the impact of bandwidth allocation
across different paths using the ECMP routing algorithm. The findings indicate that
the bandwidth proportion assigned to each path significantly influences flow completion
times and overall performance, providing guidelines for optimizing traffic engineering
in multi-path WANs. The contributions include: (1) an evaluation of BBR and Cubic
performance in DIS networks, (2) analysis of the effects of key network parameters on
transfer efficiency, (3) a study of multi-path routing strategies in sliced WANs, and (4)
practical recommendations for optimizing data transport in DIS systems. These findings are
relevant for scientists, network engineers, and service providers aiming for high-performance
networks.
The study demonstrates the potential of the FABRIC Testbed as a robust tool for
reproducible experiments and innovation in networking, particularly in high-volume data
transfer applications. It also points to future research directions, such as evaluating newer
algorithms like BBRv3 and QUIC, and leveraging advanced FABRIC resources, including
SmartNICs, FPGAs, and GPUs, to enhance performance and programmability further.
