A Statistical Network Calculus for Computer Networks

1. A Statistical Network Calculus for Computer Networks Jorg Liebeherr Department of Computer Science University of Virginia

2. Collaborators • Almut Burchard • Robert Boorstyn • Chaiwat Oottamakorn • Stephen Patek • Chengzhi Li • Florin Ciucu

3. Papers • R. Boorstyn, A. Burchard, J. Liebeherr, C. Oottamakorn. “Statistical Service Assurances for Packet Scheduling Algorithms”, IEEE Journal on Selected Areas in Communications. Special Issue on Internet QoS, Vol. 18, No. 12, pp. 2651-2664, December 2000. • A. Burchard, J. Liebeherr, and S. D. Patek. “A Calculus for End–to–end Statistical Service Guarantees.” (2nd revised version), Technical Report CS-2001-19, May 2002. • J. Liebeherr, A. Burchard, and S. D. Patek , “Statistical Per-Flow Service Bounds in a Network with Aggregate Provisioning”, Infocom 2003. • C. Li, A. Burchard, J. Liebeherr, “Calculus with Effective Bandwidth”, Technical Report CS-2003-20, November 2003. • F. Ciucu, A. Burchard, J. Liebeherr, ",A Network Service Curve Approach for the Stochastic Analysis of Networks”, ACM Sigmetrics 2005, to appear.

4. “Toy Models” in Computer Networking • Learn from Physics: Wide use of toy models … that capture key characteristics of studied system … that permit back-of-the-envelope calculations … that are usable by non-theorists • Simple models have played a major role in the evolution and development of data networks • Queueing Networks • Effective Bandwidth • (Deterministic) Network Calculus

5. (Product Form) Queueing Networks • Jackson (50’s), Kelly, BCMP (70’s) • Flow of “jobs” in system of queues and servers • Applications: Provided motivation for packet-switching (Kleinrock’s PhD thesis) Main result: Steady state probability of queue occupanceyn = (n1, n2, … , nk) : P(n ) = P(n1) P(n2) … P(nk) Limitations: • Limited to Poisson traffic • Limited scheduling algorithms

6. Effective Bandwidth Hui, Mitra, Kelly (90s) • Describes bandwidth needs of complex traffic by a number • Application: admission control in ATM networks Peak rate effectivebandwidth Mean rate Can consider: • service guarantees • wide variety of traffic (incl. LRD)  statistical multiplexing Limitations:  not well suited for scheduling

7. S3 S1 S2 Receiver Sender Snet Network Calculus • Cruz, Chang, LeBoudec (90’s) • Worst case delay and backlog bounds for fluid flow traffic • Application: design of new schedulers (WFQ) new services (IntServ). • Main result: If S1, S2 and S3 describes the service at each node, then Snet = S1 * S2 * S3describes the service given by the network as a whole. Limitations: • No random losses • No statistical multiplexing, therefore pessimistic

8. State-of-the-art • No analysis methodology is widely used today. • Today, a lot of networking research relies on simulation and measurements to validate new designs • Simulation and measurement are generally not suitable for evaluation of radically new designs

9. Motivation: Develop network calculus into new“Toy Model” Today, fundamental progress in networking is hampered by the lack of methods to evaluate how radically new designs will perform. • Opportunity: Simple (`toy') models that permit fast (`back-of-the-envelope') evaluations can become an enabling factor for breakthrough changes in networking research • Approach: Probabilistic version of network calculus (stochastic network calculus) is a candidate for a new class of toy models for networking

10. Deterministic network calculusCruz `91 Effective bandwidth in network calculusChang `94 Effective Bandwidth: J. Hui ’88Guerin et.al. ’91Kelly `91Gibbens, Hunt `91 (min,+) algebra for det. networks: Agrawal et.al. `99Chang `98LeBoudec `98 ServiceCurvesCruz `95 • Our goals: • Maintain elegance of deterministic calculus • Exploit statistical multiplexing • Try to express other models Cruz calculus with probabilistic trafficKurose `92 Exponentially/stochasti-cally. bounded burstinessYaron/Sidi `93Starobinski/Sidi `99 RateVarianceEnvelopeKnightly `97 Stochastically bounded service curveQiu et.al.`99 Related Work (small subset) 2005 1985 1990 1995 2000

11. Multiplexing Gain Multiplexing gain is the raison d’être for packet networks. Sources of multiplexing gain: • Traffic characterization and conditioning • Scheduling • Statistical Multiplexing

12. Traffic Conditioning Traffic Conditioning • Traffic conditioning is typically done at the network edge • Reshaping traffic increases delays and/or losses

13. Scheduling • Scheduling algorithm determines the order in which traffic is transmitted • Examples: • Different loss priorities  priority scheduling • Traffic with rate guarantees  rate-based scheduling (WFQ, WRR) • Delay constraints  deadline-based scheduling (EDF)

14. Worst-casebacklog Backlog Backlog Multiplexing Gain Without statistical multiplexing Worstcasearrivals Flow 1 Flow 2 Flow 3 Time With statistical multiplexing Arrivals Flow 1 Flow 2 Flow 3 Time Backlog

15. Example of Statistical Multiplexing: Retirement Savings Life expectancy: Normal(m=75, s=10) years Retiring Age: 65 years Interest: 0% Withdrawal: $50,000 per year How much money does a person need to save (with confidence of 95% or 99%)? Life expectancy in a group of N people is Normal(m, s / N). N=1 person (Individual Savings): 95% confidence: 10 + 2s = 30 years  $1.5 Mio.99% confidence: 10 + 2s = 40 years  $2 Mio. N=100 people (Pooled Savings): 95% confidence: 10 + 2s = 12 years  $600,00099% confidence: 10 + 2s = 13 years  $650,000

16. The importance of Statistical Multiplexing • At high data rates, statistical multiplexing gain dominates the effects of scheduling and traffic characterization

17. Traffic Characterization • Arrivalsfrom a flow j are a random process • Stationarity: The are stationary random processes • Independence: The and are stochastically independent

18. Regulated Arrivals Flow 1 . . . C Flow N Each flow isregulated Buffer with Scheduler Regulated arrivals Traffic is constrained by a subadditive deterministic envelope such that Leaky Buckets:

19. Effective envelope Define a function that bounds traffic with high probability  “Effective Envelope” Definition:Effective envelope for is a function such that Note: Effective envelope is not a sample path bound. Often, we need a stronger version of the effective envelope!

20. Stronger effective envelope At most one sample path is violated Deterministic envelope Never violated Samplepaths Effective envelope At any time, at most one sample path is violated Sample Paths and Envelopes Note: All envelopes are non-random functions

21. Probabilistic Sample Path Bound A strong effective envelope for an interval of length is a function which satisfies Relationship between the envelopes is established as follows: with

22. Aggregating Arrivals Flow 1 . . . C Flow N Traffic Conditioning Buffer with Scheduler Regulated arrivals Arrivals from multiple flows: Deterministic Network Calculus: Worst-case of multiple flows is sum of the worst-case of each flow

23. Effective Envelopes for aggregated flows Stochastic Calculus: Exploit independence and extract statistical multiplexing gain by calculating • For example, using the Chernoff Bound, we can obtain

24. Effective vs. Deterministic Envelope Envelopes Type 1 flows: P =1.5 Mbps r = .15 Mbps s =95400 bits Type 2 flows: P = 6 Mbps r = .15 Mbps s = 10345 bits strong effective envelopes Type 1 flows

25. Effective vs. Deterministic Envelope Envelopes Traffic rate at t = 50 msType 1 flows

26. Deterministic Service Never a delay bound violation if: Statistical Service Delay bound violation with if: Scheduling Algorithms • Work-conserving scheduler with unit rate that serves Q classes • Class-q traffic has delay bound dq • Scheduling algorithm: . . . Scheduler • Static Priority (SP): • Earliest Deadline First (EDF):

27. Statistical multiplexing makes a big difference Scheduling has small impact Statistical Multiplexing vs. Scheduling Example: MPEG videos with delay constraints at C= 622 Mbps Deterministic service vs. statistical service (e = 10-6) dterminator=100 ms dlamb=10 ms Thick lines: EDF SchedulingDashed lines: SP scheduling

28. More interesting traffic types • So far: Traffic of each flow was regulated • Next: Consider different traffic types: • On-Off traffic • Fraction Brownian Motion (FBM) traffic • Approach: Exploit literature on Effective Bandwidth • Describes traffic in terms of a function • Expressions have been derived for many traffic types

29. Effective Envelopes and Effective Bandwidth Effective Bandwidth (Kelly 1996) Given , an effective envelope is given by

30. Effective Envelopes and Effective Bandwidth Comparisons of statistical service guarantees for different schedulers and traffic types Schedulers: SP- Static PriorityEDF – Earliest Deadline FirstGPS – Generalized Processor Sharing Traffic: Regulated – leaky bucketOn-Off – On-off sourceFBM – Fractional Brownian Motion C= 100 Mbps, e = 10-6

31. Statistical Network Calculus with Min-Plus Algebra D(t) A(t) S(t)

32. Convolution and Deconvolution operators • Convolution operation: • Deconvolution operation

33. Deterministic (min,+)Network Calculus Cruz `95:A service curve for a flow is a function S such that: (min,+) results(Cruz, Chang, LeBoudec) • Output Envelope: is an envelope for the departures • Backlog bound: is an upper bound for the backlog • Delay bound: An upper bound for the delay is

34. Stochast Network Calculus An effective service curve for a flow is a function such that: (min,+) results • Output Envelope: is an envelope for the departures with probability e • Backlog bound: is an upper bound for the backlog with probability e • Delay bound: An upper bound for the delay with probability eis

35. Sender Receiver Statistical Per-Flow Service Bounds Allocated capacity C • Given: • Service guarantee to aggregate (C ) is known • Total Traffic is known • What is a lower bound on the service seen by a single flow?

36. Sender Receiver Statistical Per-Flow Service Bounds Allocated capacity C Can show: is an effective service curve for a flow where is a strong effective envelope and is a probabilistic bound on the busy period

37. Number of flows that can be admitted Type 1 flows: Goal: probabilisticdelay bound d=10ms

38. Snet Network Service Curves S3 S1 Receiver S2 Sender Deterministic Network Service Curve (Cruz, Chang, LeBoudec): If are service curves for a flow at nodes, then Snet = S1 * S2 * S3 is a service curve for the entire network.

39. Unfortunately, this network service is not very useful! Finding a suitable network service curve has been a longstanding open problem. A solution is presented in an upcoming ACM Sigmetrics 05 paper. Network Service Curve in a Stochastic Calculus Network Service Curve: If S1,, S2 , … SH , are effective service curves for a flow, then for all .

40. Effective Network Service Curve • Revise the definition of the effective service curve to • Define Theorem: A network service curve is given by with where are free parameters

41. Application of Network Service Curve • Analyze end-to-end delay of through flows for Markov Modulated On-Off Traffic • Compare delay with network service curve to a summation of per-node bounds

42. Example • C = 100 Mbos • Cross traffic = through traffic • e = 10-9 • Peak rate: P = 1.5 MbpsAverage rate: r = 0.15 Mbps • T= 1/m + 1/l = 10 msec • Addition of per-node bounds grows O(H3) • Network service curve bounds grow O(H log H)

43. Conclusions • Presented aspects of stochastic network calculus • Preserves much (but not all) of the deterministic calculus • Can express many existing results on: • Deterministic calculus • Effective bandwidth • Other models (EBB, not shown) • Many open issues

44. Conclusions