Re: [EToSat] about LOOPS(Localized Optimization of Path Segments) and etosat

[Carsten Bormann <[e-mail]>]

On Jan 3, 2019, at 18:48, Border, John <[e-mail]> wrote:
> 
> Clearly LOOPS should not focus only on satellite.  But, I think LOOPS is directly relevant to ETOSAT.  Techniques like LOOPS are likely to be a key part of the solution to the performance over satellite problem.
> 
> I will be sending out an email to ETOSAT in the next couple of days to try to kickstart discussion overall.

Here is my attempt at kickstarting:

Traditional approaches at enhancing end-to-end performance by deploying nodes on the way (performance enhancing proxies, PEPs) have employed heavy interactions with the end-to-end transport protocols, usually without their explicit consent.  With end-to-end encryption going under the transport (QUIC), this approach is no longer useful.  This is the underlying theme of the ETOSAT mailing list, but also an important motivator for the more general LOOPS approach.  (Please use the slides previously sent as a reference to node terminology.)

So what kind of performance enhancement can we do even without aggressively reaching into transport layers?

* Active queue management: AQM is certainly a good idea.  A potential benefit of LOOPS for AQM is that it can run measurements between sub-path ingress and sub-path egress nodes, as well as between overlay ingress and overlay egress.  So more data can be made available inside the system of LOOPS nodes, and this may help in active queue management.

* Loss recovery.  Packets can be recovered between sub-path ingress and sub-path egress nodes.  This can be based on local retransmissions and/or on adding and using FEC.  By cooperation of the sub-path ingress and egress nodes, the parameters for these (RTO, FEC rate, choice between retransmission and FEC) can be chosen wisely.  Overlay ingress and egress nodes can provide additional information, such as end-to-end RTT estimates, which might also help in choosing these parameters and avoiding continuing with sub-path retransmission when an end-to-end retransmission is likely to kick in.  The total load on the network decreases with replacing end-to-end by subpath retransmissions/FEC.  Tail-loss detection can benefit from mixing multiple end-to-end flows into one error-controlled subpath flow.

* Multipathing.  Again, both the overlay ingress/egress nodes and the sub-path ingress/egress nodes can help with detecting multiple paths and using them in the best possible way.

Challenges, and potential mitigations by employing advances in end-to-end transports, include:

* Reflecting congestion information up from the sub-paths to the end-to-end paths.  Sub-path ingress/egress nodes may employ additional congestion indicators such as sojourn time to sort loss events into congestion or non-congestion losses.

* Reordering may reduce end-to-end transport performance.  RACK may help us here, as may cwnd undo on spurious retransmission detection in current TCPs.  Or we could use resequencing in egress nodes, but this is generally not a favored approach.

* MTU.

Traditionally, satellite PEPs have treated the whole satellite section (uplink, downlink) as a single sub-path (hop) for optimization.  LOOPS would benefit from an active LOOPS node on the satellite.  Is that something that is on the table these days?

What we could design here includes:

* Sub-path “transport-layer style“ protocols, for retransmission and FEC.  This includes designing/selecting good FEC schemes for sub-path usage, as well as some basic encapsulation formats.
* Formats for making measurement information available between the nodes.
* Formats for making control information available between ingress and egress nodes.

Grüße, Carsten