Connection( less & oriented)
- 1. Internetworking connectionless and connection-oriented networks Malathi Veeraraghavan Mark Karol Polytechnic University Bell Laboratories mv@poly.edu mk@lucent.com Outline: » Why internetwork? » Prior work » Our proposal 6/1/99 1
- 2. Why internetwork? Connectionless (CL) Network CL Network Connection-Oriented (CO) Network Router Endpoint Switch Endpoint 6/1/99 2
- 3. Problem Statement • Applications at endpoints start sending data without warning in connectionless networks • CO networks need a connection setup phase • So how do the gateways cope with the traffic arriving from the CL networks without time to set up a connection? Networking modes Switching modes Connectionless Connection-oriented Packet-switching IP ATM Telephony network, Circuit-switching SONET/SDH, WDM 6/1/99 3
- 4. Use provisioned connections • Use provisioned connections through CO network – Suitable for some cases CL Network CL Network CO Network Provisioned connections: set up a priori based on anticipated traffic Switched connections: set up on demand as traffic arrives 6/1/99 4
- 5. Switched connections • Need switched connections for some cases – CL applications have an application-level handshake that can be used to trigger connection setups • e.g., interconnecting an Internet telephony PC to a telephone • e.g., H.245 signaling to Q.931 signaling through the PSTN phone Router CL Network Endpoint Gateway Switch CO Network Endpoint 6/1/99 5
- 6. Prior work • Interesting case - Case 3 – A choice exists of which network to use • Existing solutions: – MPOA (Multi-Protocol Over ATM) – MPLS (Multi-Protocol Label Switching) CL Network CO Network 6/1/99 6
- 7. Solutions - MPOA • MPOA: – Overlay model – Routing data not shared – Good solution if choice to use CO network made based on application needs (e.g., interactive sessions with long holding times) CL Network IP packet 10 7 5 1 1 Interactive application 1 1 (long-lived flow; SETUP if flow classifier is set to use CO network for CO Network this flow type) 6/1/99 7
- 8. Solutions - MPOA • MPOA: – Not a good solution if either CL or CO network can be used for a given application (e.g., large bulk-data transfers) CL Network IP packet 10 7 5 1 1 1 6 1 IP packet IP packet 1 CO Network 1 1 If flow classification does not detect this as a flow to be handled by the CO network, it will not take advantage of shorter path through the CO network 6/1/99 8
- 9. Solutions - MPLS • MPLS: – Peer model – Routing data is shared – Requires every CO switch to also be a CL router – Same example as last slide - large bulk-data transfer that could go either way CL Network IP packet 10 7 5 1 1 1 6 Gateway will select IP packet IP packet 1 1 CO network because SETUP SETUP path is shorter CO /CL Network Packets will be forwarded in 1 1 CL mode while IP packet IP packet SETUP SETUP connection is being set up 6/1/99 9
- 10. Proposed solution • Peer model • Routing data is shared – How is this done: routing-related actions • But, not all nodes in the CO network need to have CL capability • Problem created: – Data arrives from the CL endpoints into the gateway before connections are set up – User-plane actions 6/1/99 10
- 11. Routing related actions • Gateways running OSPF connected by a CO network (non-broadcast network) announce point-to-point links between gateways GW1 S4 R6 S2 R3 R1 GW2 S1 S5 R5 R2 R7 GW3 R4 S3 CO Network CL Network Note: switches have no CL capability 6/1/99 11
- 12. Routing related actions • Topological view of each router and gateway 2 GW1 1 5 R6 R3 1 1 2 GW2 1 R1 Shortest path from 3 1 1 1 R4 to R7 is via 1 R5 GW3 and GW2 R2 2 4 R7 GW3 R4 1 CL Network User data packets from R4 to R7 arrive at GW3 even before connection is set up 6/1/99 12
- 13. User-plane actions • IP datagrams arrive at the gateway to be carried through the CO network when no connection exists through it. – IP datagram could be carrying a TCP segment – IP datagram could be carrying a UDP datagram • CO network used only for flows classified as needing connections or those that can be handled on either network 6/1/99 13
- 14. For flows for which the CO network is to be used • TCP segment – If it is a SYN segment, hold it up, set up connection • SYN-related time-outs are large (5 sec) – If it is a data segment, then send zero-window- size acknowledgment to halt data • if persist timers get routed through some other path and new data packets arrive before the connection is set up, send another zero-window-size acknowledgment 6/1/99 14
- 15. For flows for which the CO network is to be used • UDP datagram – For applications with user-level message exchange, hold up such messages and set up connection (e.g., H.245 open logical channel) – For applications without such exchanges • use source routing to override default routes • use small-bandwidth provisioned pipes 6/1/99 15
- 16. Applications Interactive Streaming Bulk-data e.g., live or stored e.g., telnet, rlogin, e.g. ftp, smtp, http audio or video telephony Packet-switched CO networks Circuit-switched (CO) networks Small amounts of Large amounts of data transfer data transfer CL (packet-switched) networks Circuit-switched or CL networks Peer model needed for this case
- 17. Comparison of CO network options • Circuit switches – IP traffic is bursty by the time it reaches gateway owing to TCP congestion control – Circuit switching not efficient for bursty traffic • ATM switches – 20% overhead due to 10% cell header overhead + TCP acks not fitting in one cell • Switched IP connections: – Reserve bandwidth and buffer for specific flow (hard state) – No additional overhead IP (network-layer) rides over DLL 6/1/99 17
- 18. Switched IP connections • New IP routers capable of performing multi-tuple route lookups/scheduling at wire-speed – destination and source addresses – destination and source ports – protocol type and TOS (Type of Service) • Question: Are there any conditions under which a network of ATM switches or circuit switches can perform better than these “IP switches?” 6/1/99 18
- 19. Options • Option 1: – Use protocol conversion not protocol encapsulation • Avoids having to carry TCP ACKs in CO network • Much simpler transport-layer protocol can be used in CO network since the network nodes now maintain state and perform congestion control (instead of state information being maintained at endpoints) • Option 2: – Generate traffic at endpoints in mode appropriate for network used 6/1/99 19
- 20. Option 1: Protocol conversion APP APP APP APP TCP/UDP TCP/UDP AAL5 AAL5 TCP/UDP TCP/UDP IP IP IP ATM ATM ATM IP IP DLL DLL DLL DLL DLL DLL DLL DLL DLL DLL PHY PHY PHY PHY DLL PHY PHY PHY DLL PHY Endpoint Router Gateway ATM Switch Gateway Endpoint • Drawback: TCP state information about many connections needs to be held at the gateways • Feasibility as yet untested. 6/1/99 20
- 21. Option 2: Download software to endpoints CO interface Web browser CO interface Web server program Both Windows program CGI and Solaris allow for device driver addition CO device driver TCP/IP CO device driver TCP/IP Link-layer module Link-layer module CL Network Link-layer Link-layer mux/demux mux/demux CO Network 6/1/99 21
- 22. Conclusions • For applications whose data can be carried in either the CL network or CO network, internetworking should allow for the exchange of routing information (peer model) • Requiring all CO nodes to have CL capability seems too constraining (an MPLS requirement) • Hence, our proposed solution: – Share routing data – “Halt” or “turn back traffic” while setting up connections • To overcome overheads of protocol encapsulation – Perform protocol conversion, or – Download software to endpoints for CO service 6/1/99 22