Baptiste Jonglez - End-to-end mechanisms to improve latency in communication networks

The network technologies that underpin the Internet have evolvedsignificantly over the last decades, but one aspect of network performancehas remained relatively unchanged: latency.  In 25 years, the typicalcapacity or "bandwidth" of transmission technologies has increased by 5orders of magnitude, while latency has barely improved by an order ofmagnitude.  Indeed, there are hard limits on latency, such as thepropagation delay which remains ultimately bounded by the speed of light.

This diverging evolution between capacity and latency is having a profoundimpact on protocol design and performance, especially in the area oftransport protocols.  It indirectly caused the Bufferbloat problem,whereby router buffers are persistently full, increasing latency evenmore.  In addition, the requirements of end-users have changed, and theyexpect applications to be much more reactive.  As a result, new techniquesare needed to reduce the latency experienced by end-hosts.

This thesis aims at reducing the experienced latency by using end-to-endmechanisms, as opposed to "infrastructure" mechanisms.  Two end-to-endmechanisms are proposed.  The first is to multiplex several messages ordata flows into a single persistent connection.  This allows bettermeasurements of network conditions (latency, packet loss); this, in turn,enables better adaptation such as faster retransmission.  I applied thistechnique to DNS messages, where I show that it significantly improvesend-to-end latency in case of packet loss.  However, depending on thetransport 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 (suchas Wi-Fi, wired Ethernet, 4G).  The idea is to use low-latency paths forlatency-sensitive network traffic, while bulk traffic can still exploitthe aggregated capacity of all paths.  This idea was partially realized byMultipath TCP, but it lacks support for multiplexing.  Adding multiplexingallows data flows to cooperate and ensures that the scheduler has bettervisibility on the needs of individual data flows.  This effectivelyamounts to a scheduling problem that was identified only very recently inthe literature as "stream-aware multipath scheduling".  My firstcontribution is to model this scheduling problem.  As a secondcontribution, I proposed a new stream-aware multipath scheduler, SRPT-ECF,that improves the performance of small flows without impacting largerflows.  This scheduler could be implemented as part of a MPQUIC (MultipathQUIC) implementation.  More generally, these results open newopportunities for cooperation between flows, with applications such asimproving WAN aggregation.