QualityofService Qo S over BestEffort Networks I E
Quality-of-Service (Qo. S) over Best-Effort Networks I E Logical FIFO B I E E I David Harrison, Yong Xia, Omesh Tickoo, Shivkumar Kalyanaraman : “shiv rpi” Rensselaer Polytechnic Institute Shivkumar Kalyanaraman 1 Sponsors: DARPA-NMS, NSF, Intel, AT&T Labs Research
Our Goals… q HISTORY: TCP offers reliability over unreliable networks q Qo. S over best-effort (I. e. non-Qo. S) networks? ? q Building blocks for … q … performance expectations … over (I. e. as an overlay) q q …multi-domain inter-networks … q . . . that do not guarantee performance… q (I. e. best-effort or FIFO-queued networks) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 2
Talk Overview Part 0: What is Qo. S? Where are we today? Part 1: I Expected Qo. S Using Closed-loop Techniques Logical FIFO Rensselaer Polytechnic Institute B I [Eg: MSN/Yahoo + Akamai] Part 2: E I E 2 E Expected Qo. S Pipe Abstraction: H 2 H or Enterprise Multimedia Apps E E Shivkumar Kalyanaraman 3
What is Quality-of-Service (Qo. S)? q “Better performance” as described by a set of parameters or measured by a set of metrics. q Generic parameters: q Bandwidth q Delay, Delay-jitter q Packet loss rate (or loss probability) q Qo. S as perceived by the application q Transport/Application-specific parameters: q Timeouts q Percentage of “important” packets lost Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 4
What is Qo. S (contd) ? q These parameters can be measured at several granularities: q “micro” flow, aggregate flow, population. q Qo. S considered “better” if q q a) more parameters can be specified a priori b) Qo. S can be specified at a finer granularity. q Qo. S vs Co. S: Co. S maps micro-flows to classes and may perform optional resource reservation per-class q Qo. S spectrum: Best Effort Leased Line Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 5
Qo. S: Fundamental Problems Scheduling Discipline FIFO B q B In a FIFO service discipline, the performance assigned to one flow is convoluted with the arrivals of packets from all other flows! q Cannot get Qo. S with a “free-for-all” q Need to use new scheduling disciplines which provide “differentiation” or “isolation” of performance from arrival rates of background traffic Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 6
Qo. S: Fundamental Problems q Conservation Law (Kleinrock): (i)Wq(i) = K q Irrespective of scheduling discipline chosen: q Average backlog (delay) is constant q Average bandwidth is constant q Zero-sum game => need to “set -aside” resources for premium services Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 7
Qo. S Big Picture: Control/Data Planes Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 8
Eg: Integrated Services (Int. Serv) Based upon Parekh-Gallagher theorem. q Shapers@ edge, WFQ per-flow schedulers @ core, Signaling/admission control q Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 9
Trends: Differentiated Services Model Interior Router Ingress Edge Router q q Egress Edge Router Edge routers: traffic conditioning (policing, marking, dropping), SLA negotiation Interior routers: traffic classification and forwarding (near stateless core!) Qo. S complexity moving to the edges… Why not move complexity further back into edge/end-systems? q End-systems/Enterprises can then take charge of their Qo. S provisioning Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 10
Inter-domain Qo. S: Challenge Provide high quality service across ISPs q Problem: intermediate ISPs don’t have incentive to provide “good” service q e. g. , “hot-potato” routing policy q ISP 2 ? ISP 3 ISP 1 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 11
Limits: Internet Path Congestion limits E 2 E bandwidth on single e 2 e path Internet Performance Server Access Link Client Access Link Performance Saturation (even w/ many flows/path) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute Access 12 Link Speed
What we Want… Aggregate Access Capacity Performance Server Access Link(s) Rensselaer Polytechnic Institute Internet Aggregate Access capacity E 2 E Broadband Virtual Pipe Abstraction!! Performance Scaling Access 13 Link Speed Client Access Link(s) Qo. S becomes a function of aggregate access capacity & ability to recruit intermediate forwarding nodes Shivkumar Kalyanaraman
Degrees of Freedom: End-to-end, Edge-to-Edge Video Coding, Error Concealment, Unequal Error Protection (UEP) Packetization, Marking, playout Buffer Management Video Coding, Error Concealment, Unequal Error Protection (UEP) Packetization, Marking, Source Buffer Management Congestion control, Multipath, Multiflow Internet Closed-loop control Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 14
Multi-paths using Peers or Overlay Nodes Overlays or peers can provide path diversity even if multi-paths not available natively! Issue: diversity of performance (b/w, delay, loss), possible correlations… Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 15
Talk Overview Part 0: What is Qo. S? Where are we today? Part 1: Expected Qo. S Using Closed-loop Techniques [1 -path] [Eg: MSN/Yahoo + Akamai] Part 2: I Logical FIFO B I I E 2 E Expected Qo. S Pipe Abstraction: H 2 H or Enterprise Multimedia Apps [multi-paths/flows] E E E Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 16
Part 1: Overlay Closed-loop Qo. S FIFO Priority/WFQ B q B Scheduler: differentiates service on a packet-by-packet basis Loops: differentiate service on an RTT-by-RTT basis using edge-based policy configuration. q. Differentiation/Isolation meaningful in steady state only… q Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 17
Architectural Advantages of Closed Loops q Traffic management consolidated at edges (placement of functions in line with E 2 E principle) Bottleneck queue Edge system End system q Architectural Potential: q Edge-based (distributed) Qo. S services, q Edge plays in application-level Qo. S Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 18
Expected Min Rate (EMR) Service: Sample Steady State Behavior Flow 1 with 4 Mbps assured + 3 Mbps best effort Flow 2 with 3 Mbps best effort Issues: Transients, steady state oscillations/tracking errors Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 19
Expected Min Rate & Graceful Degradation No oversubscription Loop 02, 03 Web Moderate oversubscription Gross oversubscription <m 00=30, A<m stop sending 00=60 KB> Shivkumar Kalyanaraman <m 02=50, A 02=60 KB> 01=35, A 01=75 KB> Rensselaer Polytechnic Institute <m 03=10, A 03=15 KB> 20
Talk Overview Part 0: What is Qo. S? Where are we today? Part 1: Expected Qo. S Using Closed-loop Techniques I Logical FIFO Rensselaer Polytechnic Institute B I [Eg: MSN/Yahoo + Akamai] Part 2: E I E E E 2 E Expected Qo. S Pipe Abstraction: H 2 H or Enterprise Multimedia Apps [Eg: MSN/Yahoo/Google/AOL or 21 Enterprises] Shivkumar Kalyanaraman
Part 2: Best-Effort Path Multiplicity Phone modem USB/802. 11 a/b 802. 11 a Wi. Fi (802. 11 b) Ethernet Firewire/802. 11 a/b ISP-1 AS 1 Internet . . . ISP-n Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 22
Part 2: Multiplicity: Flows, Paths, Exits/Interfaces Multiple E 2 E Flows Multi-homed interfaces & ISP/AS peers Multi-paths within a routing domain Multi-AS-paths across domains W/o specifying intra-domain paths Multiple exits from an AS w/o. To specifying paths BANANAS: A Single Abstract Framework Exploitintra-domain These Shivkumar Kalyanaraman Forms Institute of Multiplicity In the Internet and Future Networks Rensselaer Polytechnic 23
Single path issues: capacity, delay, loss… High Delay/Jitter Low Capacity Lossy Network paths usually have: • low e 2 e capacity, • high latencies and • high/variable loss rates. Rensselaer Polytechnic Institute 24 Time Shivkumar Kalyanaraman
Multimedia Apps: How to Leverage E 2 E Performance Diversity? Low Perceived Loss High Perceived Capacity Low Perceived Delay/Jitter Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 25
Motivation & Key Idea q q q Motivation: Video over best-effort networks (wired/wireless) q Broadband + Multi-Homed => more access bandwidth q Goal: Virtual extension of broadband aggregate access pipe E 2 E using multi-paths: q HOME-to-HOME (H 2 H) multimedia or q Enterprise DVD/HD-quality multimedia Path Diversity: dimensions q Aggregate Capacity q Delay diversity q Loss diversity q Correlations in path performance characteristics Key: Match inherent content diversity to path diversity Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 26
SMCA: Framework Content Delay Diversity Unit Loss Diversity Unit Network Receive Buffer Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 27
SMCA: Delay Diversity Unit High Delay RANK Low Delay RANK Application Data Paths Ranked by Latency Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 28
SMCA: Delay Diversity Unit High Delay RANK Low Delay RANK Application Data Paths Ranked by Latency Early deadline packets mapped to low-delay paths Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 29
SMCA: Delay Diversity Unit High Delay RANK Transmit Queue Paths Ranked by Latency Low Delay RANK Early deadline packets (in order of rank) mapped to low-delay paths (in order of rank) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 30
SMCA: Delay Diversity Unit High Delay RANK Low Delay RANK Transmit Queue Paths Ranked by Latency Late deadline packets mapped to high-delay paths… Note: these packets leave the sender roughly at the same time as the early-deadline packets Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 31
SMCA: Delay Diversity Loss Diversity High Delay Low Delay Transmit Queue Paths Ranked by Latency Consider a delay-based group of paths and the associated packets… Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 32
SMCA: Delay Diversity Loss Diversity High Delay Low Delay Transmit Queue Paths Ranked by Latency Consider a delay-based group of paths and the associated packets… Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 33
SMCA: Loss Diversity Unit High Loss RANK n GOPs Low Loss RANK Paths Ranked by Loss Rate Re-rank Paths within this group based upon packet loss rates Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 34
SMCA: Loss Diversity Unit P B B I High Loss RANK Low Loss RANK n GOPs Paths Ranked by Loss Rate Enlarged View of Packets (with content labels) and Paths Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 35
SMCA: Loss Diversity Unit P B B I High Loss RANK Low Loss RANK n GOPs Paths Ranked by Loss Rate Map high priority packets (eg: I-frame packets) to low loss rate rank paths Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 36
SMCA: Loss Diversity Unit P B B I High Loss RANK Low Loss RANK n GOPs Paths Ranked by Loss Rate Continue map packets to low loss rank paths based upon priority (Eg: P-frames get the next set of loss-ranked paths) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 37
SMCA: Loss Diversity Unit P B B I High Loss RANK Low Loss RANK n GOPs Paths Ranked by Loss Rate Lowest priority packets get high loss rate paths (within the delay-based group of paths) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 38
SMCA: Loss Diversity Unit + FEC P-FEC I-FEC P B B I High Loss RANK Low Loss RANK n GOPs Paths Ranked by Loss Rate FEC (unequal FEC) for a GOP mapped within the same delay-group, but mapped to the higher loss paths Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 39
SMCA: Performance with increasing number of Paths Content Source Content Sink Background traffic generator Background traffic sink Num. Of Paths PSNR (d. B) 1 2 3 4 5 20. 98 22. 48 25. 42 26. 02 28. 04 Table 1. Average PSNR Variation with Number of Paths Performance with Path Diversity + Matching Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 40
Video Results: Original Video Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 41
Video Results: SMCA vs Simple Mapping Simplistic Mapping SMCA Multiplexing Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 42
SMCA gains with delay diversity Avg. Delay (ms) 300 SMCA PSNR(d. B) PT PSNR(d. B) OPMS PSNR(d. B) 21. 78 18. 73 11. 03 100 25. 12 24. 21 19. 19 50 28. 32 29. 46 24. 33 30 30. 12 31. 63 27. 96 Table 3. Gains with Delay Variation Better comparative perf. when average delay and jitter is high Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 43
SMCA gains with loss diversity Avg. Loss Prob. SMCA PSNR(d. B) PT PSNR(d. B) OPMS PSNR(d. B) 0. 4 22. 78 20. 31 11. 64 0. 35 26. 32 26. 86 18. 21 0. 1 29. 03 29. 02 24. 43 0. 05 29. 32 31. 82 26. 06 Table 2. Gains with Loss Variation Better comparative perf. when avg loss and loss variation is high Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 44
Part 2: Summary q Multi-path performance diversity is a fundamental efficiency gain yet to be leveraged on the Internet! q q q Ideas: q q q This can be achieved end-to-end! Key: must be mapped to content diversity Map late deadline packets to high latency paths Map higher priority packets to lower loss rate paths (within a delaybased group of paths) q FEC packets sent on paths different from that of associated content (FEC: lower priority) Our scheme can scale to handle lots of paths q q Possible with p 2 p networks (eg: 10 -100 kbps from single path, but 10 s of paths) Does not require MD coding, or high complexity optimization Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 45
Perspective: Traditional Qo. S: Control/Data Planes Our Overlay Qo. S BBlocks: primarily in the data-plane. Note: Skype, Kazaa/Napster/DHTs are control-plane ideas that can be combined w/ such overlay Qo. S Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 46
Perspective: Overlay Qo. S: > best-effort Qo. S spectrum … Best Effort Over. Qo. S, Overlay DPS/CSFQ, Diff -Serv Overlay Qo. S: Qo. S w/ Closed Loop * Performance Expectations (not Control Leased Line Int-Serv guarantees), * Edge/end-systems take charge of Qo. S just like reliability/availability * Perceived Qo. S Engineering by diversity matching Traffic Shortest Path … MPLS BANANAS-TE Kalyanaraman Signaled. Shivkumar TE Rensselaer Polytechnic Institute 47
Thanks ! Papers, PPTs, Audio talks: : “shiv rpi” Ps: VIDEOS of all my networking courses available freely at the above web site. pps: Audio classes on South Indian Music also available Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 48
Extra/Backup Slides Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 49
Method: Non-concave User Utility Functions q q A single user with a minimum rate Qo. S expectation (gracefully degrading into a weighted service) can be modeled with a non-concave utility function. But this kind of U-function cannot be plugged directly into Kelly’s nonlinear optimization formulation! log(xs) log(xs-0. 4) 10 log(xs) expected minimum =0. 4 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 50
Luckily, the Sum of Non-Concave U-fns is not what we want to Optimize! q U(z) = log(z-0. 6) if z > 0. 6 (expected minimum rate) 10 logz if z <= 0. 6 (graceful degradation to weighted svc) Two maxima Desired Allocation (oversubscribed) not any of the 2 maxima! • Can use strictly concave functions and define multiple optimization problems for the same Qo. S problem & • Dynamically choose a different optimization problem when oversubscribed Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 51
Handling both under- and over-subscription… • For ai, xi: (primary problem) • For aip, xip: (auxiliary problem) Weighted fairness Effective when: ai = Ai Exp. Min Rate: Effective when: If under-subscribed, solve the aux-problem; and the ai < Ai primary problem is automatically solved (note: aip = constant) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 52
Expected Min Rate (EMR) Service Building Block q Accumulation di Accumulation limit q Target Control Law Estimated accumulation in network Rensselaer Polytechnic Institute 53 Shivkumar Kalyanaraman
Range of Expected Minimum Rates Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 54
Queue Lengths Non-AQM: Q-length w/ virtual accumulation support More details: Y. Xia et al, “Accumulation-based Congestion Control, ” to appear in IEEE/ACM Transactions of Networking, early 2005. Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 55
- Slides: 55