Preliminaries: EE807 Software-defined Networked Computing KyoungSoo Park Department of Electrical Engineering KAIST

value in arriving packet’s header routing algorithm local forwarding table header value output link Interplay Between Routing, Forwarding 2

Intradomain Routing Learning routes in an autonomous system (AS) – Also called “intraAS routing” Two representative approaches – Distance vector (or Bellman-Ford) – Link state (or Dijkstra’s) Time complexity – Per-node: O(nlogn) where n = # of nodes (routers) 3

Distance Vector Algorithm Strategy: each node exchanges its DV with its neighbor whenever link cost changes – DV contains the estimated cost to every node Dynamic programming – Min path cost (x, y) = min(link cost(x,v) + path cost(v,y)) for all neighbor v of x Implementation – Routing Information Protocol (RIP) – EIGRP (Cisco-proprietary): solves limitations of RIP 4

Link State Algorithm Strategy: flood the directly-connected link’s cost to every node – Send to all nodes, but the spread information is the local link cost Link state packet (LSP) – Contains the link cost, id of the node, sequence number, TTL, etc. Implementation – Open Shortest Path First (OSPF), Intermediate System- Intermediate System (IS-IS) 5

Interdomain Routing Intradomain routing: process of finding the least-cost path to network prefix X (in the same AS) Interdomain routing: process of finding AS- level path that reaches the destination prefix X (not in the same AS) Routing: coarse-grain path (interdomain) + fine-grain path (intradomain) 6

Border Gateway Protocol (BGP) The goal of interdomain routing – Find some loop-free path to the destination – Concerned with reachability than optimality – Concerned with the policies of ASs in the path – Finding path anywhere close to optimal is considered to be a great achievement BGP advertises complete paths as an enumerated list of ASs to reach a particular network – Called a path-vector protocol – Example: /16: – How do you detect a loop? 7

Routers

Router Functionality Control plane: run routing protocols, run software on routing processor, circuit setup – Time scale: 10ms to second Data plane: forwarding, buffering, filtering, scheduling, implemented in hardware – Time scale: nanoseconds Management plane: administrator interface, analysis, configuration (traffic engineering) – Time scale: minutes to hours 9

10 Router Architecture Overview data plane control plane

decouple control and data planes by providing open standard API Control/Data Separation 11 Borrowed from Jen Rexford’s slides

(Logically) Centralized Controller Controller Platform 12 Borrowed from Jen Rexford’s slides

Protocols  Applications Controller Platform 13 Controller Application Borrowed from Jen Rexford’s slides

Software-defined Networking Logically-centralized control plane – Why? fine-grained control of the traffic No (traditional) routing protocols – Instead, there is a centralized controller When a flow comes to a switch – The switch looks up forwarding table – If the entry is found, use it to forward packets – If not, it asks the controller to set up the route OpenFlow is widely used to implement SDN – OpenFlow != SDN 14

Middleboxes

Middlebox In-network devices that manipulate packets for purposes other than packet forwarding – Inspecting, filtering, transforming packets Examples – Network address translators (NATs), firewalls, network intrusion detection systems (NIDSes), (performance enhancement, Web, WAN-accelerating) proxies, etc. Recent trend – # of deployed middleboxes >> # of deployed routers 16

Network Functions Virtualization Motivation: difficult to manage many middlebox boxes – Each box runs different service – Configuration could be a nightmare Conceptually, you have X units of Web proxy, Y units of NIDS, Z units of firewall – X, Y, Z are dynamically adjusting to the load How to implement? – Virtualization: separate the service from physical infrastructure – Horizontal scaling (or scale out): add more nodes, install software, and turn them on Vertical scaling (or scale up)? 17