Transcript PPT Version
Triggers for Transport (TRIGTRAN) Perspective IETF 57 Alias BOF Spencer Dawkins [email protected] Carl Williams [email protected] Origins of TRIGTRAN • Two Sources – Performance Implications of Link Characteristics (PILC) – Layer-two triggers (L2triggers) Bar BoF at IETF 53 • PILC completing BCPs on improving TCP-as-it-was – Protocol changes required to move TCP forward • Allison, Spencer, Carl dreamed up TRIGTRAN – – – – “Triggers for Transports” “What do links know that transports would like to know?” Transports would figure these things out eventually… Maybe links could tell them in less than several RTTs? • Constraints (as of IETF 55 BoF) – Access link, likely wireless, no multi-homing IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 2 TRIGTRAN Functionality TRIGTRAN Initiator HOST Correspondent HOST Single IP Hop TRIGTRAN Router Transport IP Layer Subnetwork Layer IETF 57 July 2003 Arbitrary Network Topology IP Layer Subnetwork Layer TRIGTRAN Perspective - Alias BOF Transport IP Layer Subnetwork Layer 3 TRIGTRAN Functionality TRIGTRAN Initiator HOST Subnetwork Event Here Transport IP Layer Subnetwork Layer Correspondent HOST Single IP Hop IETF 57 July 2003 TRIGTRAN Router Arbitrary Network Topology IP Layer Subnetwork Layer TRIGTRAN Perspective - Alias BOF Transport IP Layer Subnetwork Layer 4 TRIGTRAN Functionality TRIGTRAN Initiator HOST Subnetwork Event Here Transport IP Layer Subnetwork Layer Correspondent HOST Single IP Hop IETF 57 July 2003 TRIGTRAN Router IP Layer Subnetwork Layer Arbitrary Network Topology Notification Transport IP Layer Subnetwork Layer Notify Transport Here TRIGTRAN Perspective - Alias BOF 5 What Happened • TRIGTRAN BoFs at IETF 55 and IETF 56 • Discussed three “notifications” – Link Up, Link Down, Packets Discarded • Strawperson framework proposal – To provide notifications from middleboxes – Allowed explicit TRIGTRAN coverage requests • Room consensus to move forward on Link Up – As end-to-end implicit notification – Draft-dawkins-trigtran-linkup-00.txt • Link Down, Packets Discarded too scary – Because notifications aren’t authenticated • Don’t even THINK about “Loss Due to Errors”! IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 6 TRIGTRAN Trust Mismatch TRIGTRAN Initiator HOST Correspondent HOST Single IP Hop TRIGTRAN Router IETF 57 July 2003 Arbitrary Network Topology TRIGTRAN Perspective - Alias BOF 7 TRIGTRAN Trust Mismatch TRIGTRAN Initiator HOST Correspondent HOST Single IP Hop TRIGTRAN Router Arbitrary Network Topology <---Trust we had---> IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 8 TRIGTRAN Trust Mismatch TRIGTRAN Initiator HOST Correspondent HOST Single IP Hop TRIGTRAN Router Arbitrary Network Topology <---Trust we had---> <---Trust we wanted---> IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 9 TRIGTRAN Trust Mismatch TRIGTRAN Initiator HOST Correspondent HOST Single IP Hop TRIGTRAN Router Arbitrary Network Topology <---Trust we had---> <---Trust we wanted---> <--------Trust we settled on for LinkUp------> IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 10 What were the implications? • There were other issues, but fundamentally … – Unauthenticated notifications = neutered responses – “Advisory notifications” – don’t ignore the ACK stream • Link Down as DoS bait – – – – Stop transmitting before transport detects loss? Based on unauthenticated notification? NO! Notification had to be advisory – add complexity SCTP interest for switchover – if they could trust it – – – – Retransmit without slowing down? Based on unauthenticated notification? NO! Notification had to be advisory – add complexity Loss Due To Errors as extreme case • Packets Discarded = loss without congestion IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 11 Lessons for Alias • TRIGTRAN agreed on one notification – LinkUp – – – – Redefined as end-to-end notification Functionality limited to short-circuiting RTO backoff Has no effect unless RTO has already taken place Slow start with cwnd=1 still happens after first ACK • Notifications complicate transport state machines – If we can’t trust them, even more complexity • If we can’t redefine a notification as end-to-end? – No support for it in TRIGTRAN, at IETF 56 BoF – Link Down is the extreme case, of course • No future for middlebox transport guidance – Without authentication IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 12 TRIGTRAN drafts • Strawperson problem statement – draft-dawkins-trigtran-probstmt-01.txt • Strawperson framework – draft-dawkins-trigtran-framework-00.txt • LinkUp specification – draft-dawkins-trigtran-linkup-00.txt IETF 57 July 2003 TRIGTRAN Perspective - Alias BOF 13