Computer Science is Largely about Abstractions

Transcription

1 Metarouting Timothy G. Griffin João Luís Sobrinho Computer Laboratory Instituto detelecomunicações University of Cambridge Instituto Superior Técnico Cambridge, UK Lisbon, Portugal SIGCOMM ugust 23, 2005

2 Computer Science is Largely about bstractions Instances Capturing (almost) all instances Parsers Yacc Data Management SQL-based systems Systems + application code Routing Protocols Metarouting (??)

3 Why do this? No one-size-fits-all IGP GP is now a widely used IGP! Hard to define, standardize, and deploy new routing protocols (or minor modifications to existing protocols) Just standardize Metarouting language and leave it up to operator community to standardize protocols using high-level specs It s fun!

4 Idea #1 Let s try something radical -- keep these separate! Protocol = Mechanism + Policy +.??? How are routing messages exchanged and propagated? (example: Link- State, Path Vector) How are adjacencies established? How are the attributes of a route described? How does configuration attach path characteristics? How are best routes selected?

5 Idea #2 Use Routing Policy lgebras as basis for Policy Component Network Routing with Path Vector Protocols: Theory and pplications João Sobrinho. SIGCOMM 2003 (to appear in ToN Oct. 2005) m + n m n Generalize Shortest Paths λ σ λ σ

6 Routing Policy lgebras How paths are transformed by application of labels L : link labels : Σ + L Σ = (Σ, <=, L,, O ) How paths are described and compared Σ : path signatures <= is a preference relation over Σ : (complete) x, y in Σ, x <= y or y <= x (or both) (transitive) x, y, z in Σ, if x <= y and y <= z then x <= z subset of signatures that can be associated with originated routes.

7 Example --- ddition (DD) 2 1 DD(3, 6) max signature max label L Σ DD(n, m) is SM if 0 < n <= m

8 Guarantees? We want protocols that are nice! always converge, for every network state unique solution (perhaps modulo some class) no forwarding loops (after convergence)

9 Correctness Monotonicity (M): σ ε Σ/, λ ε L σ <= λ Strict Monotonicity (SM): σ ε Σ/, λ ε L σ < λ σ σ Isotonicity (I): σ,β ε Σ/, λ ε L σ <= β λ σ <= λ β SM I ssoc. vectoring Link-state with generalized Dijkstra Link-state with local vector simulation

10 n algebra for OSPF? (hand-coded from careful reading of RFC 2328) ε (1, ε, σ) (1, (1, v), σ) (1, (2, v), σ) (2, ε, σ) (2, (1, v), σ) (2, (2, v), σ) (1, λ) (1, ε, λ ε) (1, ε, λ σ) (1, (1, v), λ σ) (1, (2, v), λ σ) (2, ε, λ σ) (2, (1, v), λ σ) (2, (2, v), λ σ) (1, (1, v), λ) (1, (1, v), λ ε) (1, (2, v), λ) (1, (2, v), λ ε) (2, λ) (2, ε, λ ε) (2, ε, λ σ) (2, (1, v), λ σ) (2, (2, v), λ σ) (2, (1, v), λ) (2, (1, v), λ ε) (2, (2, v), λ) (2, (2, v), λ ε) <1, > = intra-area route <2, > = inter-area route <{1,2, λ> = normal route <{1,2, <1, v>, λ> = type I external <{1,2, <2, v>, λ> = type II external

11 Routing lgebras are a good start, but The algebraic framework does not, by itself, provide a way of constructing new and complex algebras. lgebra definition is hard Proofs are tedious Modifications to an algebra s definitions are difficult to manage

12 Idea #3 Routing lgebra Meta-Language (RML) ::= (base algebras) Op() (unary operator) Op (binary operators) bstract syntax for generating new lgebras Goals Want to automatically derive properties (M, SM, ) of the algebra represented by an RML expression from properties of base algebras and preservation properties of operators Simplicity Expressiveness

13 Lexical Product Preservation properties (σ1, σ2) (λ1, λ2) (λ1 σ1, λ2 σ2) M M M SM SM (, _) = (_, ) = M SM SM Preference is Lexical order This suggests a design pattern for SM: 1 2 i (i+1) n all M SM don t care SM

14 Point-wise application? Preservation properties (σ1, σ2) SM SM M λ1 (λ1 σ1, σ2) λ2 (σ1, λ2 σ2) SM M M M SM M M M M κ() λ σ1 λ1 σ1 κ() M M = κ() κ() κ σ1 SM M

15 Scoped Product Preservation properties (σ1, σ2) SM SM SM (λ1, σ3) (λ1 σ1, σ3) SM M M λ2 (σ1, λ2 σ2) Can be used to implement IGP/EGP-like information hiding

16 Programmatic Labels = (Σ, 8, L,,0 ) prog() = (Σ, 8, L,,0 ) λ ::= λ λ1; λ2 reject if π then λ1 else λ2 λ σ = λ σ (λ1; λ2) σ = λ1 (λ2 σ) reject σ = (if π then λ1 else λ2) σ λ1 σ if π(σ) = λ2 σ o.w. prog() M M SM SM

17 MyFirstIGP prog( distance : DD(2000, 2000) router-path: SimpleSeq(1000, 30) bandwidth: WIDTH(10000) tags: TGS(string[100])) metapolicy MyIGP { programatic { lex { attribute distance { label link-weight { mode local; default: 1 data { addition ; attribute router-path { label router-id { mode local nodal; data { Sequence IPv4; This is SM attribute bandwidth { label link-bandwidth { mode local; data { WIDTH 1000; attribute tags { data { TGS string[100];

18 Using MyIGP policy my-export { metapolicy MyIGP; if (empty(router-path)) { insert tags sales center NE GP-17 { router-id ; mechanism link-state hard-state neighbor { metapolicy MyIGP; import [set link-bandwidth OC-12; my-import]; export my-export; policy my-import { metapolicy MyIGP; if ( data center in tags) { if (bandwidth < OC-12) { set link-weight 100; else { set link-weight 10 else { set link-weight 20;

19 Metarouting Specification, n RML expression Label Modalities RP = <E, M, LM> set of mechanisms that can be bound to adjacencies

20 Ongoing, Future Work RML More operators t the protocol level Inter-operation operators Implementation Using XORP ( With Mark Handley (UCL) and others Hijack GP Routing Metaprotocol

21 Disjuntion ( Injection ) Preservation properties τ τ σ1 σ2 SM SM SM λ1 λ2 λ1 σ1 λ2 σ2 SM M M M SM M M M M ι τ(σ)

22 at the protocol level? <, M1, LM1> τ <, M2, LM2> = < τ, M1 + M2, LM1 + LM2> Not sure what + means Nice way of thinking about administrative distance Perhaps OSPF is really something like <RES, Path-Vector, LM1> <DD, Link-State, LM2>

23 Migration operators