Broadcast Avoidances - Neighborhood Hash Assessment

First follow evaluations of the two scenarios described before to check if or when the neighborhood hashing approach successfully reduces broadcast overhead in these targeted cases (section 1).

Afterwards, an evaluation follows to show that we do not avoid rebroadcasts too often. That is that we do not hinder BATMAN from finding or using best paths (section 2 + 3).

1) Example Scenario Evaluation

Scenario A)

image0

Many mesh nodes wired to one, big switch

Assuming all nodes have a 1GBit/s connection to the switch. Then a node has for any neighbor:

neigh->throughput
= neigh->min_throughput
= neigh->max_throughput
= iface->min_throughput_other
= iface->max_throughput_other
= 1000MBit/s

That means the ingress TX, egress TX as well as ingress RX and egress RX check evaluate to:

fwd-penalty(1000) < 1000

Which is true as long as a hop penalty is configured (default).

This means, that a rebroadcast avoidance successfully takes place for scenario A).

Scenario B)

image1

Two mobile nodes clustering around a broadcast transmitter

Assuming all nodes have a 100MBit/s connection to each other, except the cluster of S, A1 and A2.

When S broadcasts, will A1/A2 refrain from rebroadcasting?

Neighborhood hash

First of all, even though not all neighbors see the same neighborhood - in fact, all nodes S and R1 to R5 see different neighbors, some neighbors still have the same neighborhood hash: S, A1, A2. This allows for instance A1/A2 to check whether to rebroadcast packets from S or not.

Case: S, A1, A2 have a 100MBit/s connection

Yes, see evaluation for scenario A).

Case: S, A1, A2 have a 150MBit/s connection

The evaluation will look as follows:

TX

RX

ingress

penalty(150) < 100

penalty(150) < 100

egress

penalty(150) < 100

penalty(150) < 100

Which is true for all cases with the half-duplex penalty: 75 < 100.

So yes, a rebroadcast avoidance would successfully take place here.

Case: S, A1, A2 have a 300 MBit/s connection

The evaluation will look as follows:

TX

RX

ingress

penalty(300) < 100

penalty(300) < 100

egress

penalty(300) < 100

penalty(300) < 100

So no, A1/A2 will not detect a rebroadcast avoidance possibility as the statement “150 < 100” is false.

Optimality Assessment

Introduction / Concept

Let’s start with a simple, three node example:

image2

Should B rebroadcast?

Here, A is our initial broadcaster and B the node having to decide whether or not to rebroadcast on the same, incoming interface.

To decide this, we will compare the direct path between A and C vs. the path from A to C via B:

image3

Is (AB ⊕ BC) a shortcut or a detour?

With a simple throughput metric the path from A to C via B is a detour exactly when either the connection from A to B or from B to C (or both) is a bottleneck. That is if either AB or BC offers a smaller throughput than AC.

That means B can perform an ingress check (AB < AC?) and an egress check (BC < AC?) individually.

To make sure, that a rebroadcast is unnecessary, B would need to perform such an ingress and egress check for any neighbor Cn (any neighbor other than A):

image4

If B notices that for all these neighbors Cn it is a bottleneck and no improvement then B can safely avoid a rebroadcast.

Simplification

Instead of performing all the checks described above, the modified rules are used in the implementation with the following goals:

  • Less information to exchange

  • Less computational overhead

  • Supposed to work even with large neighborhoods

Performing the checks stated above has the following disadvantages: For one thing, the computational overhead could be significant for neighborhoods of a certain size. For another, B does not even know the throughput from A to B or A to C (or C to A or C to B - we will see later why we might need those). B only knows its own TX throughput towards other neighbors and not the other way around.

Instead of every neighbor node frequently broadcasting a full list of TX values and every neighbor performing all these checks for any neighbor with any neighbor combination, the implemented neighbor hash approach applies a huge simplification needing considerably less computations and information exchange:

Instead of checking all (AB < ACn?) and (BCn < ACn?) combinations, it tries to perform “one check that rules them all” (with the downside of maybe not catching all potential rebroadcast avoidance cases).

Proofs

We need to ensure that avoiding a rebroadcast does not break our optimality promises. That is that we do not break that a route eventually converges to its optimum path.

While the assessment above showed when optimality is not broken, the following mathematic assessment shows that our slightly modified rules still hold optimality.

Proof: Ingress TX (bcast)

We want to proof:

  • p(Amax) < Amin => p(AB) < ACn for any Cn.

One has:

  • a ≤ b => p(a) ≤ p(b)

  • AB ≤ Amax

  • Amin ≤ ACn

AB ≤ Amax
=> p(AB) ≤ p(Amax)

p(Amax) < Amin
=> (p(AB) ≤ p(Amax)) < Amin
=> (p(AB) ≤ p(Amax)) < (Amin ≤ ACn)
=> p(AB) ≤ p(Amax) < Amin ≤ ACn
=> p(AB) < ACn

Proof: Egress TX (bcast)

We want to proof:

  • p(Bmax) < Amin => p(BCn) < ACn

One has:

  • a ≤ b => p(a) ≤ p(b)

  • BCn ≤ Bmax

  • Amin ≤ ACn

BCn ≤ Bmax
=> p(BCn) ≤ p(Bmax)

p(Bmax) < Amin
=> (p(BCn) ≤ p(Bmax)) < Amin
=> (p(BCn) ≤ p(Bmax)) < (Amin ≤ ACn)
=> p(BCn) ≤ p(Bmax) < Amin ≤ ACn
=> p(BCn) < ACn

Proof: Ingress RX (OGM2)

We want to proof:

  • p(AB) < MIN(Amin, Cn-min) => p(AB) < CnA for any Cn.

One has:

  • MIN(a, b) ≤ b

  • Cn-min ≤ CnA

p(AB) < MIN(Amin, Cn-min)
=> p(AB) < (MIN(Amin, Cn-min) ≤ Cn-min)
=> p(AB) < (MIN(Amin, Cn-min) ≤ (Cn-min ≤ CnA))
=> p(AB) < MIN(Amin, Cn-min) ≤ Cn-min ≤ CnA
=> p(AB) < CnA

Proof: Egress RX (OGM2)

We want to proof:

  • p(MAX(Amax, Cn-max)) < MIN(Amin, Cn-min) => p(AB) < CnA for any Cn.

One has:

  • a ≤ b => p(a) ≤ p(b)

  • MIN(a, b) ≤ b

  • a ≤ MAX(a, b)

  • AB ≤ Amax

  • Cn-min ≤ CnA

Amax ≤ MAX(Amax, Cn-max)
=> (AB ≤ Amax) ≤ MAX(Amax, Cn-max)
=> AB ≤ MAX(Amax, Cn-max)
=> p(AB) ≤ p(MAX(Amax, Cn-max))

p(MAX(Amax, Cn-max)) < MIN(Amin, Cn-min)
=> (p(AB) ≤ p(MAX(Amax, Cn-max)) < MIN(Amin, Cn-min)
=> (p(AB) ≤ p(MAX(Amax, Cn-max)) < (MIN(Amin, Cn-min) ≤ Cn-min)
=> (p(AB) ≤ p(MAX(Amax, Cn-max)) < (MIN(Amin, Cn-min) ≤ (Cn-min ≤ CnA))
=> p(AB) ≤ p(MAX(Amax, Cn-max) < MIN(Amin, Cn-min) ≤ Cn-min ≤ CnA
=> p(AB) < CnA

Hash Collision Probability

sha512 provides 256 bits of security against collision attacks. With 4.8e29 inputs, that is with 4.8e29 different neighborhoods, we would have a collision probability of 1e-18.

This number of inputs would be reached if over 500 years a new neighborhood were formed 3e10 times per nanosecond.

See also: