E-Locus - Institutional Repository of the University of Crete - Implementation of the link monitor for a minimal-state max-fair rate regulation mechanism

Home Implementation of the link monitor for a minimal-state max-fair rate regulation mechanism

Results - Details

[Add to Basket]

Identifier

000413381

Title

Implementation of the link monitor for a minimal-state max-fair rate regulation mechanism

Alternative Title

Υλοποίηση του ελεγκτή συνδέσμων για έναν μηχανισμό ελάχιστης κατάστασης κατανομής μεγίστου-ελαχίστου ρυθμού ροών

Author

Γιαννόπουλος, Δημήτριος Μ.

Thesis advisor

Κατεβαίνης, Μανόλης
Χρυσός, Νικόλαος

Reviewer

Μπίλας, Άγγελος
Φατούρου, Παναγιώτα

Abstract

Ιn cloud computing, a large number of servers within the same facility work collectively in order to complete challenging tasks. Current research examines ways to replace the expensive and high-energy processors of today with cheaper, cost-efficient ARM-based processor clusters. These environments will ultimately consist of many lower-capacity end-nodes, potentially increasing the intensity of inter-server communication. At the same time, VM migration, checkpointing and storage contribute to sudden traffic bursts, which are responsible for transient congestion phenomena. In this thesis, we develop an implementation for a novel congestion control scheme that can replace the TCP congestion control for RDMA transfers. Our scheme can (i) throttle the offensive flows, (ii) keep the backlogs outside of the network, and (iii) allocate max-min-fair rates to flows. The protocol does not require per-flow state inside the network, and relies on simple components that can readily be implemented on network interfaces and on switch outputs. Furthermore this thesis covers more than the implementation of the congestion control scheme, despite the fact that the concept was given complete. Due to great adversity that was encountered during this implementation, a lot of research had to be conducted, and this thesis covers solving the problems that surfaced That being said most of the goals have been met and what is left is well defined for future work, the important components are ready for placement, and have displayed promising results, both in quickly resolving contention, before allowing congestion to manifest, and in general scalability. The simulation results demonstrate that the scheme reacts promptly to congestion, while the implementation results, having made tests on an UltraScale+ FPGA, prove that the scheme is scalable as the components only require about 1% of the FPGA, with over 10 instances of the component placed.

Language

English

Subject

Network congestion

Patcket switching

Μέγιστοι - ελάχιστοι ρυθμοί

Μεταγωγή πακέτων

Συμφώρηση δικτύων

Issue date

2017-11-24

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses