Bridging centralized programming and distributed control planes Ryan Beckett Ratul Mahajan Todd Millstein Jitu Padhye David Walker
Hypothesis on smart solutions in networks Simple Overprovision, best effort, Present solutions (eng + ops + $/bit) optimize mean perf, relaxed (distributed) control, …. Past Total cost Smart Intelligent resource allocation, multiple priorities, optimize tail perf, tight (centralized) control, … solutions Infrastructure size or complexity
Network programming journey Distributed Centralized programming programming + Resilience Distributed control plane − Programmability Centralized Control plane
Programming (configuring) networks is error-prone
Programming (configuring) networks is error-prone 50-80% of outages are the 60% of network downtime is result of human error caused by human error -Yankee group 2002 -Juniper 2008
Network programming journey Distributed Centralized programming programming + Resilience + Resilience Distributed control plane − Programmability + Programmability − Resilience Centralized Control plane + Programmability Network verification
Programming distributed control planes is hard Network-wide policies Router-level mechanisms Prefer one neighbor over another Set consistent, per-link preferences Don’t use my network as transit Tag incoming routing info Keep traffic within a region Program import and export filters Aggregate prefixes externally based on various route attributes Prov Cust Peer R3 R2 R4 R1
Propane: Centrally programming distributed control planes A language for expressing A compiler that configures of network-level objectives router-level mechanisms • Path constraints and relative • Configurations are policy- preferences (fallbacks) compliant under all failures
Example #1: A backbone network block “Peer” block “Peer” tag “Peer” tag “Peer” Peer2 Peer1 R1 LP=99 R2 LP=99 LP=100 LP=101 filter 16.4.0.0/16 filter 16.4.0.0/16 Cust Goals define notransit = { true => not transit ({Peer1, Peer2})} • No transit between peers define preference = { true => exit (R2>R1>{Peer1, Peer2})} • Prefer R2 > R1 > Peer{1,2} define ownership = {16.4.0.0/16 => end (Cust)} • Limit Cust to 16.4.0.0/16 define main = notransit and preference and ownership
Example #2: A data center network Goals Backbone • Keep local prefixes internal • Aggregate global prefixes as PG X Y Y Attempt #1 C C D D G G H H • Don’t export from G, H to external A A B B E F E F • Aggregate externally as PG PG1 PG2 PL1 PL2 Global Local Services Services
Example #2: A data center network Goals Backbone • Keep local prefixes internal • Aggregate global prefixes as PG X Y Y Attempt #1 C C D D G G H H • Don’t export from G, H to external A A B B E F E F • Aggregate externally as PG PG1 PG2 PL1 PL2 Global Local Services Services
Example #2: A data center network Goals Backbone • Keep local prefixes internal PG • Aggregate global prefixes as PG X Y Y Attempt #2 C C D D G G H H • Don’t export from G, H to external A A B B E F E F • Aggregate externally as PG PG1 PG2 PL1 PL2 • Valley-free routing Global Local Services Services
Example #2: A data center network Goals Backbone • Keep local prefixes internal PG • Aggregate global prefixes as PG X Y Y Attempt #2 C C D D G G H H • Don’t export from G, H to external A A B B E F E F • Aggregate externally as PG PG1 PG2 PL1 PL2 • Valley-free routing Global Local Services Services
Example #2: A data center network Goals Backbone • Keep local prefixes internal PG • Aggregate global prefixes as PG X Y Y PG1 blackholed! Attempt #2 C C D D G G H H • Don’t export from G, H to external A A B B E F E F • Aggregate externally as PG PG1 PG2 PL1 PL2 • X, Y block routes through the other Global Local Services Services
Example #2: A data center network Goals Backbone • Keep local prefixes internal • Aggregate global prefixes as PG X Y Y define ownership = {PG1 => end (A), C C D D G G H H PG2 => end (B), PL1 => end (E), A A B B E F E F PL2 => end (F)} define locality = { {PL1, PL2} => always (in)} PG1 PG2 PL1 PL2 control { aggregate (PG , in -> out )} Global Local define main = routing and locality Services Services
Propane compiler Propane Front End Constraint Language Regular Expression-based IR Regular IR Topology Product graph Failure analyses Abstract BGP Vendor-independent BGP Cisco Vendor configurations Juniper
Propane Propane Regular IR Regular Step 1: Combine modular constraints IR Prefix-by-prefix intersection of constraints define ownership = {PG1 => end (A), PG2 => end (B), PG1 => end (A) PL1 => end (E), PG2 => end (B) PL2 => end (F)} PL1 => always(in) and end (E) define locality = { {PL1, PL2} => always (in)} PL2 => always ( in ) and end (F) control { aggregate (PG , in -> out )} define main = routing and locality
Propane Propane Regular IR Step 2: Expand constraints in to regular expressions Regular any = out*.in+.out* IR end (X) = (Σ*.X) always (X) = (X)* (out*.in*.(X ∩ in).out+)| exit (X) = (out*.in+.(X ∩ out).out*) start (X) = (X.Σ*) avoid (X) = (!X)* waypoint (X) = (Σ*.X.Σ*) Step 3: Reduced syntax true => A.(X >> Y). out * true => (A.X. out *) >> (A.Y. out *)
Regular PG construction: An Example IR Product graph Y X W A D B C E Z Policy: (W.A.C.D. out ) >> (W.B. in +. out )
Regular PG construction: An Example IR Product graph Y X W A D B C E Z Policy: (W.A.C.D. out ) >> (W.B. in +. out )
PG construction: Reversed policy automata Y X W A D B C E Z out D C A W 5 0 1 2 3 4 (W.A.C.D. out ) ACDE B in out W B (W.B. in +. out ) 0 1 2 3 4 ACDE
PG construction: Graph generation start Y X W (Z,1,1) (Y,1,1) (X,1,1) (W,1,1) A D B (B,-,2) (E,-,2) (D,2,2) (A,-,2) C E (C,3,2) (C,-,2) (D,-,2) Z (W.A.C.D. out ) >> (W.B. in +. out ) out D C A W (B,-,3) (A,4,2) 0 1 2 3 4 5 (W,-,4) (W,5,-) ACDE B in W out {2} {1} B 0 1 2 3 4 end ACDE
PG construction: Graph generation start Y X W (Z,1,1) (Y,1,1) (X,1,1) (W,1,1) A D B (B,-,2) (E,-,2) (D,2,2) (A,-,2) C E (C,3,2) (C,-,2) (D,-,2) Z (W.A.C.D. out ) >> (W.B. in +. out ) out D C A W (B,-,3) (A,4,2) 0 1 2 3 4 5 (W,-,4) (W,5,-) ACDE B in W out {2} {1} B 0 1 2 3 4 end ACDE
PG construction: minimization (loop analysis) start start (Z,1,1) (Y,1,1) (X,1,1) (W,1,1) (Z,1,1) (Y,1,1) (X,1,1) (B,-,2) (E,-,2) (D,2,2) (A,-,2) (E,-,2) (D,2,2) (C,3,2) (C,-,2) (C,3,2) (C,-,2) (D,-,2) (B,-,3) (B,-,3) (A,4,2) (A,4,2) (W,-,4) (W,-,4) (W,5,-) (W,5,-) {2} {1} {2} {1} end end
Product Compilation to ABGP graph Abstract start BGP (Z,1,1) (Y,1,1) (X,1,1) Idea 1: Restrict advertisements to PG edges (E,-,2) (D,2,2) • Encode PG state in community tag • Incoming edges — import filters (C,3,2) (C,-,2) • Outgoing edges — export filters (B,-,3) (A,4,2) Let BGP find some allowed path dynamically (W,-,4) (W,5,-) {2} {1} end
Product Compilation to BGP graph Abstract start BGP (Z,1,1) (Y,1,1) (X,1,1) (E,-,2) (D,2,2) C allows import from D with tag (2,2) (C,3,2) (C,-,2) C exports to A,B with tag (3,2) (B,-,3) (A,4,2) (W,-,4) (W,5,-) {2} {1} end
Compilation to BGP start A better path exists in the (Z,1,1) (Y,1,1) (X,1,1) network, but is not used! (E,-,2) (D,2,2) Y X W (C,3,2) (C,-,2) A D B C E (B,-,3) (A,4,2) (W,-,4) (W,5,-) Z (W.A.C.D. out ) >> (W.B. in +. out ) {2} {1} end
Compilation to BGP start Idea 2: Synthesize local preferences (Z,1,1) (Y,1,1) (X,1,1) • Direct BGP towards best path (E,-,2) (D,2,2) • Under all combinations of failures (C,3,2) (C,-,2) (B,-,3) (A,4,2) Let BGP find the best allowed path dynamically (W,-,4) (W,5,-) {2} {1} end
Compilation to BGP Router A start match peer=C comm=(3,2) export peer←W, comm←(4,2), comm← noexport, MED←80 (Z,1,1) (Y,1,1) (X,1,1) Router B match peer=C (E,-,2) (D,2,2) export peer←W, comm←( -,3), comm←noexport, MED←81 (C,3,2) (C,-,2) Router C match[lp=99] peer=E, comm=(-,2) export peer←B, comm←( -,2) (B,-,3) (A,4,2) match[lp=100] peer=D, comm=(2,2) export peer←A,B, comm←(3,2) (W,-,4) (W,5,-) Router D match regex=(X + Y) {2} {1} export peer←C, comm←(2,2) end …
Compilation to BGP Router A start match peer=C comm=(3,2) export peer←W, comm←(4,2), comm← noexport, MED←80 (Z,1,1) (Y,1,1) (X,1,1) Router B match peer=C (E,-,2) (D,2,2) export peer←W, comm←( -,3), comm←noexport, MED←81 (C,3,2) (C,-,2) Router C match[lp=99] peer=E, comm=(-,2) export peer←B, comm←( -,2) (B,-,3) (A,4,2) match[lp=100] peer=D, comm=(2,2) export peer←A,B, comm←(3,2) (W,-,4) (W,5,-) Router D match regex=(X + Y) {2} {1} export peer←C, comm←(2,2) end …
Recommend
More recommend
