policing congestion response in an internetwork using re-feedback Bob Briscoe 1,2 Arnaud Jacquet 1 , Carla Di Cairano- Gilfedder 1 , Alessandro Salvatori 1,3 , Andrea Soppera 1 & Martin Koyabe 1 1 BT Research, 2 UCL, 3 Eurécom
intro intro intro intro the problem: policing congestion response • host response to congestion: voluntary the idea the idea the idea • short and long term congestion • short: policing TCP-friendliness (or any agreed response) • long: policing file-sharing (selfish), zombie hosts (malicious/careless) incentives incentives incentives access capacity rate eg. TCP rate cumulative inverse deployment deployment deployment flows T √ρ prop’nal response path congestion, ρ • network policing users’ congestion response: voluntary generalise generalise generalise • a network doesn’t care if users cause congestion in other networks 2
intro intro intro intro very serious problem the idea the idea the idea • a few unresponsive (UDP) flows wasn’t a problem • converged IP network • initially ~30-50% of bits inelastic (mostly voice), for BT • internetwork similar incentives incentives incentives • can’t police required response to path congestion, if you don’t know it • each element only sees local congestion • network can’t reliably see e2e feedback (IPsec encryption, lying, route asymmetry) • can’t hope inelastic apps ask to be unresponsive (Diffserv/signalling) deployment deployment deployment • because those that don’t ask can free-ride anyway • due to lack of evidence of their ‘crime’ • capacity investment risk unacceptable if can’t prevent free-riding • uncontrollable demand dynamics and suppressed incentive to supply generalise generalise generalise • risk of repeated congestion collapse (alarmist?) 3
intro intro intro intro previous work the idea the idea the idea • detect high absolute rate [commercial boxes] • sampled rate response to local congestion [RED + sin bin] incentives incentives incentives • transport control embedded in network [ATM] • honest senders police feedback from rcvrs [ECN nonce] deployment deployment deployment generalise generalise generalise 4
intro intro intro intro wouldn’t it be nice if... ...we can: our approach • the big idea #1 the idea the idea the idea • then 2 sub-ideas based on... • source declared downstream path characteristics to network • network economics & incentives • rational networks (not users) • no fiddling with user pricing incentives incentives incentives • everyone was truthful: • challenge: break and improve – endpoints and networks • incremental deployment idea #4 • around unmodified IP routers deployment deployment deployment • BUT limited header bits slows • deployment could be incremental attack detection considerably • generalisations • QoS • we could solve more general • DoS mitigation generalise generalise generalise Internet Architecture problems • flow start incentives – capacity allocn & accountability [NewArch] • inter-domain traffic engineering • non-IP internetworks 5
intro intro intro path characterisation via data headers state of the art TTL 255 the idea the idea the idea the idea ECN marking rate 152 0.7% 0.5% 105 incentives incentives incentives resource index 0 0 resource index along path along path deployment R 1 deployment deployment N B S 1 N A N D generalise generalise generalise 6
242 intro intro 255 intro 250 249 242 S 1 0 1 R 1 N 1 N 2 5 7 N 5 1 245 252 the idea the idea the idea the idea 2 S 2 0 N 4 1 N 3 3 R 2 255 before re-feedback 254 245 incentives incentives incentives target at destination standardised to 16, say – 242 + 255 + 16 29 24 23 16 S 1 0 1 R 1 N 1 N 2 5 7 N 5 deployment deployment deployment 15 1 22 16 23 2 S 2 0 N 4 2 N 3 3 + 255 + 16 R 2 26 24 generalise generalise generalise 26 after re-feedback 27 – 245 25 15 downstream knowledge upstream 7
intro intro intro downstream path characterisation TTL 255 the idea the idea the idea the idea ECN rate 152 0.7% 0.5% 105 incentives incentives incentives resource index 0.1% 0 0 resource index along path along path re-TTL 166 deployment R 1 deployment deployment N B S 1 N A N D 119 re-ECN 16 0 0 generalise generalise generalise -0.5% -0.6% -0.7% 8
intro intro intro incentives: preamble • so far, policing relies on self-incrimination?... the idea the idea the idea • focus initially on congestion • header processing not just additive/subtractive • generalises to monotonic functions (eg combinatorial probability of ECN marking) incentives incentives incentives incentives • downstream unloaded delay (~TTL/2) has identical incentive properties • to aid understanding • solely graphical visualisation (see paper for maths) • imagine that header carries a real number • normalise: monotonically decreasing to target at zero deployment downstream deployment deployment path metric ρ i 0 generalise generalise generalise resource index along path, i 9
intro intro intro incentive framework: user-network the idea the idea the idea downstream policer incentivises understatement path metric, ρ i dropper incentivises overstatement incentives incentives incentives incentives i policer dropper deployment deployment deployment Snd Rcv generalise generalise generalise 10
generalise deployment incentives incentives the idea intro generalise deployment incentives incentives the idea the idea intro intro generalise deployment 11 dropper egress probability distribution downstream congestion Rcv dropper i 0 downstream dropper path metric statement statement naïve at rcvr, ρ n
statement statement intro intro intro dropper i Rcv the idea the idea the idea penalising uncertain misbehaviour incentives incentives incentives incentives 1 idea #2 systematic adaptive cheating, drop ∆ ρ nc deployment deployment deployment ∆ ∆ ∆ ∆ ρ nc probability stateless downstream congestion dropper probability distribution truncated/dropped downstream generalise generalise generalise if signature prevalent in discards path metric spawn focused dropper(s) at rcvr, ρ n 0 12
statement statement intro intro intro dropper i Rcv the idea the idea the idea if everyone honest minimise false incentives incentives incentives incentives positives no systematic cheating, ∆ ρ nc = 0 deployment deployment deployment stateless downstream congestion dropper probability distribution adaptive ∆ ∆ ρ nc ∆ ∆ downstream drop generalise generalise generalise path metric probability at rcvr, ρ n 0 13
intro intro intro typical dropper simulation (note log scale) -0.3 -0.2 -0.1 0 0.1 0.2 0.3 1 the idea the idea the idea truncated 0.1 honest traffic unaffected 0.01 penalty prob. 0.001 incentives incentives incentives 1e-04 1e-05 1e-05 deployment deployment deployment 1e-04 0.001 0.01 0.1 dishonest traffic generalise generalise generalise 1 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 14
intro intro intro congestion, flow delay, … policer each packet header carries the idea the idea the idea prediction of its own downstream path eg. TCP rate flow idea #3 policer TCP- incentives incentives incentives incentives friendly downstream congestion, ρ i check/enforce agreed congestion response deployment deployment deployment policer Snd generalise generalise generalise ρ i path metric, downstream 1 5
also bounded flow state policer implemented - using sampling intro intro intro ingress TCP policer: stateful implementation unloaded delay, ρ 1,1 ρ 1,1 ρ 1,1 ρ 2,1 ρ 2,1 congestion, ρ 2,1 the idea the idea the idea packet size, s s s downstream metrics ∆ t in packet headers path congestion ≈ downstr congestion at internetwork ingress incentives incentives incentives incentives ≈ ≈ ρ 2,1 ≈ ≈ p ≈ upstr RTT + 2 * downstr delay path RTT ≈ ≈ ≈ ≈ T 0 + 2 ρ 1,1 T deployment deployment deployment s 3 ≈ x TCP T 2 p generalise generalise generalise x = s/ ∆ t 1 6
intro intro intro incentive compatibility – hosts �ρ 0c the idea the idea the idea net value to 0 both end-points, � U dominant strategy incentives scheduler incentives incentives incentives dropper /policer dropper push-back ideal R 1 S 1 deployment deployment deployment overstatement of downstream path • incentivise: practical metric at source generalise generalise generalise �ρ 0c • responsible actions 0 • honest words 1 7
generalise deployment incentives incentives the idea intro generalise deployment incentives incentives the idea the idea intro intro generalise deployment policer 1 8 incentive framework Snd ρ i path metric, downstream congestion pricing routing dropper i Rcv
Recommend
More recommend
