AUKERA BERDINTASUNA SUSTATZEA FOMENTAR LA IGUALDAD DE OPORTUNIDADES

In this section, we introduce the concept of spurious OVER sessions. Also, we show how their side effects can invalidate simple assumptions that apparently hold in any iBGP topology, and have been used in previous research work.

Definition 5.1. Given an iBGP topologyB, an OVER session x ↔ y is spuri- ous if one of the two routers is not in the top layer, i.e., ifx 6∈ TBory 6∈ TB. Spurious sessions are not frequent in today’s ISP networks. Vendor guidelines also suggest to not deploy them [156]. Nevertheless, spurious sessions have been proposed to solve visibility issues [104, 105], and pre-

spurious OVERs will appear in any ISP which defines a full-mesh of iBGP sessions among the clients of a route-reflector (see [193, 179] for exam- ples). Another scenario in which spurious OVERs are likely to appear is iBGP reconfigurations. For example, current best practices to replace an iBGP full-mesh with route reflection [68] suggest to progressively intro- duce UP sessions before removing the full-mesh. Hence, OVER sessions initially in the full-mesh are likely to become spurious in intermediate con- figurations. Similar considerations hold when reconfiguring from route- reflection to full-mesh (e.g., to improve path visibility for load-balancing purposes). As illustration, 100 out of the 120 possible (83%) per-session or- derings create dissemination anomaly when reconfiguring the OVER-RIDE gadget without the initial spurious OVER between a and e (Fig. 5.5) to a full-mesh.

Route dissemination deceptions

As discussed in Section 5.4, the OVER-RIDE provides an example of how a spurious OVER improves egress point visibility at some routers, but po- tentially worsens visibility at other routers. In the gadget, the side effect of adding a spurious OVER is counter-intuitive because it induces a change in the route dissemination process at routerr without affecting the egress point selected by r. This contradicts the intuition that a connected iBGP topology guarantees that every router eventually learns at least one route for any given prefix.

Unfortunately, some previous works are based on that intuition. In par- ticular, [104, 105] assume that adding an OVER session can only improve route visibility, while [17, 18] assume that a route-reflector r can “hide” a route to a neighboring router v only if it has a closer alternative egress point.

More generally, spurious OVER sessions show that the concept of valid signaling path is not a good abstraction to study the ability of a router to learn a route to a given prefix. In order to better understand this property, we introduce the concept of dissemination correctness.

Definition 5.2. LetB be a signaling correct iBGP topology. Then, B is dis- semination correct if all the routers inB are guaranteed to receive at least one route to prefix p in the stable state, for any non-empty set of egress points forp.

Dissemination correctness does not depend on interdomain routing nor on the set of egress points currently learning routes for given pre- fixes. That is, it is a topological property. Dissemination correctness dif- fers from both signaling and forwarding correctness. Indeed, a signaling correct topology is not guaranteed to be dissemination correct. As an illus- tration, the OVER-RIDE gadget is signaling correct, but not dissemination correct. Also, a dissemination correct topology is not guaranteed to be forwarding correct. An an illustration, the network depicted in Fig. 5.3 is dissemination correct (all routers receive a route forp1), but not forward-

ing correct. The three properties actually complement each other: signal- ing correctness deals with routing anomalies that can prevent BGP from converging; dissemination correctness deals with issues in the route prop- agation process; forwarding correctness deals with forwarding anomalies caused by the interaction between iBGP and IGP.

Signaling and forwarding correctness deceptions

Besides affecting dissemination correctness, a single spurious OVER can prevent an iBGP topology to be either signaling or forwarding correct.

e0 u0 _e 2 u1 u0u2e2* u0x e0* u0u1e0 u1e0* u1u2e2* p1 p1 x u2 u0u1u2e2 u1u0x e0 _u 2e2* u2u0x e0* u2u1u0x e0 x e0* . . . x e2* . . .

Figure 5.6 Spurious OVER can create routing oscillations.

Consider Fig. 5.6. Every router is equipped with a list of valid signal- ing paths, sorted in decreasing order of preference. Observe that(u1, e0)

is a spurious OVER session. We now show that iBGP cannot converge in this configuration. Assume by contradiction that a stable state exists, and consider the routing choice at routeru2. Sinceu2receives a route directly

frome2, it is not possible thatu2 does not select any route for prefixp1.

Hence, we have the following cases.

• u2steadily selects(u2e2). In this case, u1will use its most preferred

path(u1 u2 e2), preventing u0 from selecting (u0 u1 e0). Thus, u0

will select(u0 x e0), and eventually announce it to u2. Because of

path preferences,u2 should switch to (u2 u0 x e0), yielding a con-

tradiction.

• u2 steadily selects (u2 u1 u0 x e0). This involves that u1 steadily

selects(u1 u0x e0), leading to a contradiction, since path (u1 e0) is

always available atu1 and is more preferred than(u1u0x e0).

• u2steadily selects(u2u0 x e0). This implies that u0steadily selects

(u0 x e0), and u1 is forced to select (u1 e0), since it does not re-

ceive path (u2 e2) from u2. This leads to a contradiction, sinceu0

All the cases lead to a contradiction, hence a stable state does not exist in the topology in Fig. 5.6. Observe that the path preferences highlighted in the figure can result from the standard BGP decision process (Table 1.1) if the IGP topology is such that dist(x, e0) < dist(x, e2), dist(u0, e0) <

dist(u0, e2), dist(u2, e0) < dist(u2, e2), and dist(u1, e0) = dist(u1, e2).

In this case,x, u0, andu2 prefer paths based on the closest egress point,

while u1 prefers eBGP routes received from e2 over those received from

e0 for egress-id. Ties are broken by shorter cluster-list and lower

neighbor address criteria.

e0 e2 x p1 p1 e1 y z p1 iBGP full-mesh

Figure 5.7 Spurious OVER can create forwarding loops.

Forwarding correctness can also be affected by the presence of spuri- ous OVER sessions. Consider the topology in Fig. 5.7, and assume that x steadily selects path (x e2), while z steadily selects path (z e0), because

of the IGP distances. Since those paths are learned via an OVER session,x andz will not propagate their best route to y, hence y will be forced to select the route frome1. Ify is on x’s shortest path to e2 andx is on y’s

shortest path toe1, then a loop arises forp1.