SCION Research Questions

3.1. ISD, AS identity

• How many ASes do we expect in the SCION network model?¶

• How many ISDs? What is the ontology of an ISD? Per geographic area only?¶

• One AS belongs to many ISDs?¶

3.2. Segment Dissemination

Control servers return a large number of path segments. This can cost considerable bandwidth / network egress while at the same time overloading clients with an unnecessarily large numbers of segments, mostly consisting of redundant information in terms of duplicate link and hops.¶

• This problem may be more problematic in ASes with many end hosts (e.g. IoT), or end hosts with little computing power or little spare bandwidth.¶

• Getting a full path to a remote endhost may require three round-trips with the control server.¶

There are multiple possible and independent solution steps here:¶

• Compression (idea suggested by Francois Wirz): Segments could be stored in a way that duplicate information (hops & links) is only stored once and the segments contain only references to the hops and links.¶

• Allow queries from start AS to end AS across multiple segments. This should be very easy to implement and would be compatible with the current wire protocol (protobuf).¶

  • This would reduce the number of round trips to one.¶

  • It would reduce the number of returned segments because only segments that actually connect to other segments would need to be returned.¶

• Predefine some policies that can be resolved by the control server, e.g. ANY, BEST_LATENCY, BEST_BANDWIDTH, BEST_PRICE, BEST_CO2. For these, a control server could simply calculate 5-10 good paths and return these. Moreover these could be cached for commonly requested remote ASs. If a user requires a custom policy they can still resort to requesting actual segments.¶

Doing path computation on the control server will initially increase computational cost. However, it would substantially decrease network egress. Caching of paths should reduce CPU cost, maybe even below the current cost for retrieving a large amount of segments from the local database and sending them over the network interface.¶

Examples for requesting CORE segments between different ISDs or within an ISD (as of 2024-07-12):¶

3.3. Routing Policies and Traffic Engineering

Reduced adoption due to limited routing policy possibilities, such as a (core-)AS does not want to accept transit traffic unless it starts/ends in ASs with special properties. For example a GEANT AS does not want to allow transit traffic unless it originates or ends in another research AS.¶

One solution could be to add a “confirm full path”-flag to certain segments. If this flag is set, the full path (all segments) needs to authorized by all ASes that insist on authorizing it. This is obviously less scalable but may be viable for ASes that insist on such policies. This also allows for “secret” policies.¶

Collateral: this probably needs a data plane change. Conceptually, we have only a single resulting segment, and that segment needs to be used in full, e.g. no on-path trickery.¶

3.4. DRKey

DRKey is a key distribution system that scales well with the number of endpoints in the network. It relies on two things:¶

• Two sides of a key: A fast side, and a slow side. Sometimes called fast and slow side of the derivation. The ability of deriving a key very quickly on the fast side is necessary for most of its use cases.¶

• A grouping of endpoints (such as ASes): The pieces necessary to derive a key, namely the L1 keys, are communicated to each keystore at each grouping (e.g., a keystore per AS).¶

The questions related to DRKey are the following ones (not comprehensive):¶

• Do we want to have any possible authorization that is at the moment carried out at the data-plane, be moved to the control-plane? This could include e.g. authorization to deliver traffic depending on the source, but also things like port numbers/ranges per source, etc.¶

  • Could this obsolete firewalls? What else would be necessary?¶

  • What do we mean when we say "authorize transit"?¶

• Could perfect forward secrecy DRKey be useful, and should we research it?¶

  • What would be the trust model? Do end-users trust their ASes to ephemerally create personal keys?¶

  • What would be the attacker model?¶

  • Which use cases are relevant?¶

For more info: [I-D.garciapardo-drkey].¶

3.5. SCMP Authentication

In SCION, we would like to have SCMP (SCION Control Message Protocol) include authentication for some of the message types, e.g. the interface down type, as it would affect the path choices that the endpoint, and even the source AS, can make.¶

We propose to use DRKey as the mechanism to use to derive the authentication key, where the fast path would be on the infrastructure side (e.g. the border router in the case of an interface down type of message), and the slow side being on the intended endpoint destination for that SCMP message (e.g. the endpoint receiving the SCMP interface down message). However, we have identified a number of possible issues (not comprehensive):¶

• Denial of Service/Capability Attacks: If an endpoint receives (too) many SCMP messages, it will need (too) many resources just to authenticate their origin. Most of these messages could just be sent to the endpoint to exhaust its processing capacity.¶

• Sending an SCMP message in certain cases might be an amplification factor: If a border router sends an SCMP message (e.g. interface down) on all cases, even with small packets, there is the risk of having that border router sending a lot of traffic to a possibly unintended recipient, e.g. when the packet is not source validated.¶

3.6. Proof of transit

FABRID [KRAHENBUHL2023] and EPIC [LEGNER2020].¶

3.7. NAT

At this moment, the SCION implementation is not compatible out-of-the-box with NAT'ed devices, regardless of whether these devices are end-hosts, or even running SCION services. This is due to the (UDP-IP) underlay being modified by the NAT mechanism, but not the internal destination SCION address. Although this does not concern the SCION protocols themselves, we want to check that this will not be a problem. Critically, the SCION header needs to contain the SRC address as seen by the border router so that the border router can forward incoming response packets to the correct NAT device and port.¶

Possible solutions:¶

• With IPv6 underlay, this problem disappears. // TODO Clarify why it disappears? IS the idea that we can remove NATs if everyone would use IPv6?¶

• Introduce a mechanism so that the SCION border router can report the NATed address to an endhost (similar to a STUN server).¶

4.1. Link load balancing

Links may get overloaded because the SCION routing system fails to distribute load properly over different links. New/different links might be underutilized.¶

If links become overloaded, there are several ways to handle that. Non comprehensively:¶

• Squeeze: send an SCMP message to trigger end-hosts to use an alternative path¶

• Steer: send and SCMP to trigger users to ask CS for a better path¶

• Reduce: hand over very short lived paths, let the end-hosts wait for the path to expire so that they request new paths and (hopefully) decide on a different path.¶

• Recommend: let the end-hosts know which paths are recommended by the AS at this time.¶

If a link has good properties, many AS will disseminate segments, therefore paths that go through this link and the link may become overloaded. See Simon Scherrer's work on Braess Paradox.¶

Either there needs to be some constant control by all clients to not choose the best theoretical path, but the one that works best. Or we need to find a way that control servers do not disseminate “good” links to all end-hosts.¶

The current consensus is that end-hosts can use multi-pathing and “automatically” converge on the best path, i.e. creating an equilibrium. Again, see Simon Scherrer's work on Braess Paradox.¶

4.2. Reverse Path Refreshment

When a client contacts a server, it is usually understood that it wants the server to use the reverse path to answer back. It the server uses that path for a long period of time, the path will eventually expire. How to standardize the process of refreshment?¶

• The server must ask the CS for a path, regardless of the client's policy.¶

• The client (somehow) sends a new packet with a new path, prompting the server to use this path from now on.¶

There are some nuances: Usually the server's API will store the initial address of the client to be used through all the session. We might need to take this into account.¶

A related question: how long before expiration should we still use a path? How do we handle that?¶

Do we actually need to solve this reverse path refresh problem?¶

• CONTRA: It is probably rare that a server needs to send data for a long time without the application layer protocol requiring the client to ever answer back.¶

• PRO: The client may happen to have an old-ish path. If we can't refresh, the client always needs to consider whether a path is valid "long enough", which might only be possible to guess.¶

• CONTRA: Sending keep-alives sounds like a connection based protocol. It alo means we need to figure out when to stop sending keep alives.¶

• CONTRA: It may be better to solve this in the application layer or in the overlay protocol, where we we know more about potential length of the session, or whether this is a singular request/answer type of exchange, or whether more frequent keep-alives are anyway required.¶