–
October 23, 2020
It is my pleasure to invite you to my PhD defense whose subject is " End-to-end mechanisms to improve latency in communication networks". The defense will take place on October 23 at 2 pm, with two possible ways to attend:
- either in Salle de séminaire 1 in the IMAG building (ground floor).
- either online using this link: https://meet.univ-grenoble-alpes.fr/b/bap-937-j46
The defence will take place in English. The jury is composed of:
- M. Martin Heusse, Professor, Grenoble INP, PhD supervisor
- M. Bruno Gaujal, Directeur de recherche, Inria, PhD co-supervisor
- M. André-Luc Beylot, Professor, INP Toulouse - ENSEEIHT, Referee
- M. Guillaume Urvoy-Keller, Professor, Université Côte d'Azur, Referee
- Mme Isabelle Guérin Lassous, Professor, Université Lyon 1, Examiner
- Mme Anna Brunström, Professor, Karlstads Universitet (Sweden), Examiner
Here is the thesis abstract:
The network technologies that underpin the Internet have evolved significantly over the last decades, but one aspect of network performance has remained relatively unchanged: latency. In 25 years, the typical capacity or "bandwidth" of transmission technologies has increased by 5 orders of magnitude, while latency has barely improved by an order of magnitude. Indeed, there are hard limits on latency, such as the propagation delay which remains ultimately bounded by the speed of light. This diverging evolution between capacity and latency is having a profound impact on protocol design and performance, especially in the area of transport protocols. It indirectly caused the Bufferbloat problem, whereby router buffers are persistently full, increasing latency even more. In addition, the requirements of end-users have changed, and they expect applications to be much more reactive. As a result, new techniques are needed to reduce the latency experienced by end-hosts. This thesis aims at reducing the experienced latency by using end-to-end mechanisms, as opposed to "infrastructure" mechanisms. Two end-to-end mechanisms are proposed. The first is to multiplex several messages or data flows into a single persistent connection. This allows better measurements of network conditions (latency, packet loss); this, in turn, enables better adaptation such as faster retransmission. I applied this technique to DNS messages, where I show that it significantly improves end-to-end latency in case of packet loss. However, depending on the transport protocol used, messages can suffer from Head-of-Line blocking: this problem can be solved by using QUIC or SCTP instead of TCP. The second proposed mechanism is to exploit multiple network paths (such as Wi-Fi, wired Ethernet, 4G). The idea is to use low-latency paths for latency-sensitive network traffic, while bulk traffic can still exploit the aggregated capacity of all paths. This idea was partially realized by Multipath TCP, but it lacks support for multiplexing. Adding multiplexing allows data flows to cooperate and ensures that the scheduler has better visibility on the needs of individual data flows. This effectively amounts to a scheduling problem that was identified only very recently in the literature as "stream-aware multipath scheduling". My first contribution is to model this scheduling problem. As a second contribution, I proposed a new stream-aware multipath scheduler, SRPT-ECF, that improves the performance of small flows without impacting larger flows. This scheduler could be implemented as part of a MPQUIC (Multipath QUIC) implementation. More generally, these results open new opportunities for cooperation between flows, with applications such as improving WAN aggregation.
