Скачать презентацию Edge-based Traffic Management Building Blocks I Logical FIFO Скачать презентацию Edge-based Traffic Management Building Blocks I Logical FIFO

9ce35590087909c59e89a361538a9941.ppt

  • Количество слайдов: 44

Edge-based Traffic Management Building Blocks I Logical FIFO E B I E E I Edge-based Traffic Management Building Blocks I Logical FIFO E B I E E I David Harrison, Yong Xia, Shiv Kalyanaraman, Rensselaer Polytechnic Institute [email protected] rpi. edu http: //www. ecse. rpi. edu/Homepages/shivkuma Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 1

Overview q Private Networks vs Public Networks q Qo. S vs Congestion Control: the Overview q Private Networks vs Public Networks q Qo. S vs Congestion Control: the middle ground ? q Overlay Bandwidth Services: q Key: deployment advantages q A closed-loop Qo. S building block q Services: Better best-effort services, Assured services, Quasileased lines, App-level Qo. S… Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 2

Motivation: Site-to-Site VPN Over a Multi-Provider Internetwork Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 3 Motivation: Site-to-Site VPN Over a Multi-Provider Internetwork Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 3

Virtual ISP: Network-level Overlay q q q Avoid crossing ISP boundaries q Each ISP Virtual ISP: Network-level Overlay q q q Avoid crossing ISP boundaries q Each ISP will provide good service; V-ISP can easily verify it Allocate/buy service across each ISP and compose them Network (IP)-level overlay GPo. P (core) ISP 2 Proxy (edge) ISP 3 Proxy (edge) ISP 1 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 4

Our Model: Edge-based building blocks I Logical FIFO E B I E E I Our Model: Edge-based building blocks I Logical FIFO E B I E E I New: Closed-loop control ! Policy/ Bandwidth Broker Model: Inspired by diff-serv; Aim: further interior simplification Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 5

Closed-loop BB: Bandwidth Sharing Priority/WFQ B FIFO B Scheduler: differentiates service on a packet-bypacket Closed-loop BB: Bandwidth Sharing Priority/WFQ B FIFO B Scheduler: differentiates service on a packet-bypacket basis q Loops: differentiate service on an RTT-by-RTT basis using edge-based policy configuration. q Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 6

Queuing Behavior: Without Closed-loop Control Bottleneck queue End system Shivkumar Kalyanaraman Rensselaer Polytechnic Institute Queuing Behavior: Without Closed-loop Control Bottleneck queue End system Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 7

Queuing: With Closed Loops q q Bottleneck management issues consolidated at edges q Key: Queuing: With Closed Loops q q Bottleneck management issues consolidated at edges q Key: Transparent and lossless loop schemes Potential: q Edge-based Qo. S services, q Edge plays in application-level Qo. S, active networking. . Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 8

Closed-loop Building Block Reqts #1. Edge-to-edge overlay operation, #2. Robust stability #3. Bounded-buffer/zero-loss, #4. Closed-loop Building Block Reqts #1. Edge-to-edge overlay operation, #2. Robust stability #3. Bounded-buffer/zero-loss, #4. Minimal configuration/upgrades + incremental deployment #5. Rate-based operation: for bandwidth services Not available in any congestion control scheme… q Related work: NETBLT, TCP Vegas, Mo/Walrand, ATM Rate/Credit approaches q Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 9

Queuing at One Router: Arrival / Service Curves flow i at router j q Queuing at One Router: Arrival / Service Curves flow i at router j q arrival curve Aij(t) & service curve Sij(t) q cumulative q continuous q non-decreasing q if no loss, then q bit Aij(t) delay Sij(t) b 2 b 1 queue t 1 t 2 time Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 10

Accumulation: Series of Routers ingress 1 j j+1 J egress dj fi Λi q Accumulation: Series of Routers ingress 1 j j+1 J egress dj fi Λi q Λi, j+1 μi we have q μij define accumulation which is a time-shifted, distributed sum of buffered bits of flow i at all routers 1 through J Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 11

Accumulation (Contd) ingress 1 fi q Λi j j+1 dj μij Λi, j+1 J Accumulation (Contd) ingress 1 fi q Λi j j+1 dj μij Λi, j+1 J egress μi then 12 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 12

Accumulation vs Queuing q queue qij(t) -- num of bits of flow i queued Accumulation vs Queuing q queue qij(t) -- num of bits of flow i queued in a fifo router j q accumulation ai(t) -- num of bits of flow i queued in a set of fifo routers 1~J is the forward direction propagation delay. q the collective queuing behavior of a series of fifo routers looks similar to that of one single fifo router Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 13

Accumulation: Physical Meaning 1 j j+1 J dj fi μij Λi Λi, j+1 μi Accumulation: Physical Meaning 1 j j+1 J dj fi μij Λi Λi, j+1 μi … 1 Rensselaer Polytechnic Institute … j time 14 j+1 J Shivkumar Kalyanaraman 14

Edge-based Control (EC) policy 1 j j+1 J dj fi Λi μij Λi, j+1 Edge-based Control (EC) policy 1 j j+1 J dj fi Λi μij Λi, j+1 μi q control objective : keep q if , no way to probe increase of available bw; q control algorithm : Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 15

EC schemes monaco q accumulation estimation: out-of-band / in-band q congestion response: additive inc/additive EC schemes monaco q accumulation estimation: out-of-band / in-band q congestion response: additive inc/additive dec (aiad), etc q vegas q accumulation estimation: in-band q congestion response: additive inc / additive dec (aiad) q riviera q accumulation estimation: in-band q congestion response: additive inc / multiplicative dec using egress rate (aimd-er) q 16 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 16

Recall: accumulation theory … 1 Rensselaer Polytechnic Institute … j time 17 j+1 J Recall: accumulation theory … 1 Rensselaer Polytechnic Institute … j time 17 j+1 J Shivkumar Kalyanaraman

Accumulation vs. Monaco Estimator 1 j j+1 J dj fi μij Λi Λi, j+1 Accumulation vs. Monaco Estimator 1 j j+1 J dj fi μij Λi Λi, j+1 μi … … out-of-band in-band ctrl pkt time 1 j j+1 J Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 18

Accumulation vs. Monaco estimator 1 jf fi Λi djf μij Jb jb+1 jf+1 Jf Accumulation vs. Monaco estimator 1 jf fi Λi djf μij Jb jb+1 jf+1 Jf Λi, j+1 djb μi jb 1 data ctrl out-of-bd ctrl classifier fifo in-band ctrl, data pkt 19 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 19

ec: monaco q congestion estimation: q out-of-band q congestion and in-band control packets response: ec: monaco q congestion estimation: q out-of-band q congestion and in-band control packets response: (AIAD) q if qm < α, cwnd(k+1) = cwnd(k) + 1; q if qm > β, cwnd(k+1) = cwnd(k) – 1; [ 1 = α < β = 3 ] 20 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 20

ec: vegas cwnd congestion avoidance slow start q congestion Time estimation: q define qv ec: vegas cwnd congestion avoidance slow start q congestion Time estimation: q define qv = ( cwnd / rttp – cwnd / rtt ) * rttp; where rttp is round trip propagation delay (basertt) q congestion response: q if qv < α, cwnd(k+1) = cwnd(k) + 1; q if qv > β, cwnd(k+1) = cwnd(k) – 1; Rensselaer Polytechnic Institute 21 [1=α<β=3] Shivkumar Kalyanaraman

Vegas Accumulation Estimator q the physical meaning of qv q rtt = rttp + Vegas Accumulation Estimator q the physical meaning of qv q rtt = rttp + rttq [ rttq is queuing time ] q qv = ( cwnd / rttp – cwnd / rtt ) * rttp = ( cwnd / rtt ) * ( rtt – rttp ) = ( cwnd / rtt ) * rttq [ if rtt is typical ] = sending rate * rttq [ little’s law ] = packets backlogged [ little’s law again ] so vegas maintains α ~ β number of packets queued inside the network q it adjusts sending rate additively to achieve this q 22 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 22

Accumulation vs. Vegas estimator 1 fi q Λi jf μij Jb jb+1 jf+1 djf Accumulation vs. Vegas estimator 1 fi q Λi jf μij Jb jb+1 jf+1 djf Jf Λi, j+1 djb jb μi 1 data ack Backlogv 23 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 23

Vegas vs. Monaco estimators q Vegas accumulation estimator q ingress-based q round trip (forward Vegas vs. Monaco estimators q Vegas accumulation estimator q ingress-based q round trip (forward data path and backward ack path) q sensitive to ack path queuing delay q sensitive to round trip propagation delay measurement error q Monaco accumulation estimator q egress-based q one way (only forward data path) q insensitive to ack path queuing delay q no need to explicitly know one way propagation delay Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 24

Riviera congestion estimation: q in-band techniques, similar as vegas q congestion response: q 25 Riviera congestion estimation: q in-band techniques, similar as vegas q congestion response: q 25 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 25

Riviera: stability and fairness q lyapunov q each function flow i maximizes ( utility Riviera: stability and fairness q lyapunov q each function flow i maximizes ( utility – penalty ) q proportionally fair 26 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 26

Linear Network Topology 8 U U 8 U En 0 E 0 U Bn Linear Network Topology 8 U U 8 U En 0 E 0 U Bn E 1 U E 2 U 100 Mbps I 1 B 0 I 00 send rate (Mbps) I 10 U U 8 I 2 4 ms B 1 8 8 U E 00 I 0 U All links are 4 ms, 100 Mbps. I=ingress, E=egress, U=UDP, B=Bottleneck 27 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 27

Stability and Fairness 28 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 28 Stability and Fairness 28 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 28

Utilization 29 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 29 Utilization 29 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 29

Utilization w/ Reverse Path Congestion 30 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 30 Utilization w/ Reverse Path Congestion 30 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 30

Queue, Utilization w/ Basertt Errors Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 31 31 Queue, Utilization w/ Basertt Errors Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 31 31

Service Differentiation: Loss-based or Accumulation-based ? 32 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 32 Service Differentiation: Loss-based or Accumulation-based ? 32 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 32

Overlay Edge-to-edge Bandwidth Services q Idea: Use the EC scheme as a closed-loop building Overlay Edge-to-edge Bandwidth Services q Idea: Use the EC scheme as a closed-loop building block for a range of Qo. S services q Basic Services: no admission control q “Better” best-effort services q Denial-of-service attack isolation support q Weighted proportional/priority services q Advanced services: edge-based admission control q Assured service emulation q “Quasi-leased-line” service q Key: no upgrades; only configuration reqts… Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 33

Scalable Best-effort TCP Service Without Overlay Scheme With Overlay Scheme Queue distribution to the Scalable Best-effort TCP Service Without Overlay Scheme With Overlay Scheme Queue distribution to the edges => can manage more efficiently Co. V vs. No of Flows FRED at the core vs. FRED at the edges with overlay control between edges Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 34

Scalable Best-effort TCP Service Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 35 Scalable Best-effort TCP Service Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 35

Edge-based Isolation of Denial of Service/Flooding TCP starting at 0. 0 s UDP flood Edge-based Isolation of Denial of Service/Flooding TCP starting at 0. 0 s UDP flood starting at 5. 0 s Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 36

Edge-based Assured Service Emulation q Backoff Differentiation Policy: r= if no congestion r+D min(r, Edge-based Assured Service Emulation q Backoff Differentiation Policy: r= if no congestion r+D min(r, b. AS m, b. BE(m-a)+a) if congestion 1 > b. AS > b. BE >> 0 q q Backoff little (bas) when below assurance (a), Backoff (bas) same as best effort when above assurance (a) Backoff differentiation quicker than increase differentiation Service could be potentially oversubscribed (like frame-relay) q Unsatisfied assurances just use heavier weight. Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 37

Bandwidth Assurances Flow 1 with 4 Mbps assured + 3 Mbps best effort Flow Bandwidth Assurances Flow 1 with 4 Mbps assured + 3 Mbps best effort Flow 2 with 3 Mbps best effort Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 38

Quasi-Leased Line (QLL) q q Assume admission control and route-pinning (MPLS LSPs). Provide bandwidth Quasi-Leased Line (QLL) q q Assume admission control and route-pinning (MPLS LSPs). Provide bandwidth guarantee. Key: No delay or jitter guarantees! q Adaptation in O(RTT) timescales q Average delay can be managed by limiting total and per. VL allocations (managed delay) Policy: r= r+D if no congestion max(a, b. BE(m-a)+a) if congestion 1 > b. BE >> 0 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 39

Quasi-Leased Line Example Best-effort VL starts at t=0 and fully utilizes 100 Mbps bottleneck. Quasi-Leased Line Example Best-effort VL starts at t=0 and fully utilizes 100 Mbps bottleneck. Best-effort rate limit versus time Background QLL starts with rate 50 Mbps Best-effort VL quickly adapts to new rate. Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 40

Quasi-Leased Line Example (cont) Bottleneck queue versus time Starting QLL incurs backlog. Unlike TCP, Quasi-Leased Line Example (cont) Bottleneck queue versus time Starting QLL incurs backlog. Unlike TCP, VL traffic trunks backoff without requiring loss and without bottleneck assistance. Requires more buffers: larger max queue Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 41

Quasi-Leased Line (cont. ) Worst-case queue vs Fraction of capacity for QLLs Single bottleneck Quasi-Leased Line (cont. ) Worst-case queue vs Fraction of capacity for QLLs Single bottleneck analysis: b B/w-delay q< 1 -b products For b=. 5, q=1 bw-rtt Simulated QLL w/ edge-to-edge control. Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 42

Current Work q With bottlenecks consolidated at the edge: q What diff-serv PHBs or Current Work q With bottlenecks consolidated at the edge: q What diff-serv PHBs or remote scheduler functionalities can be emulated from the edge ? q What is the impact of congestion control properties and rate of convergence on attainable set of services ? q Areas: q Control plane architecture for large-scale overlays q Application-level Qo. S: edge-to-end problem q Dynamic (short-term) services q Congestion-sensitive pricing: congestion info at the edge q Edge-based contracting/bidding frameworks Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 43

Summary q Private Networks vs Public Networks q Qo. S vs Congestion Control vs Summary q Private Networks vs Public Networks q Qo. S vs Congestion Control vs Throwing bandwidth q Edge-based Building Blocks & Overlay services: q A closed-loop Qo. S building block: EC framework q Accumulation concept q Monaco, Vegas, Riviera Schemes: estimation issues q Basic services, advanced services Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 44