Impact of Traffic Aggregation on Network Capacity and Quality of Service

Transcription

1 Impact of Traffic Aggregation on Network Capacity and Quality of Service Towela Nyirenda-Jere Information and Telecommunications Technology Center University of Kansas 1/46

2 Outline Introduction and Motivation Network Calculus Single-Link Analysis Network Analysis Sensitivity Analysis Significance Future Work 2/46

3 Introduction and Motivation Introduction Evolution of Internet from research to commercial Growth in volume and diversity of traffic => redesign of Internet architecture => revision of engineering rules 3/46

4 Introduction and Motivation The Network Spectrum Aggregate handling + abundant capacity Low Semi-aggregate handling + simpler traffic management + moderate capacity Per-flow handling + complex traffic management + minimal capacity Increasing Network Capacity Capacity Spectrum Traffic Handling Spectrum Increasing Complexity of Traffic Handling Service Spectrum Guarantee High Guarantee 4/46

5 Introduction and Motivation The Problem given the varying levels of complexity and differing capacity requirements of aggregate, semi-aggregate and per-flow traffic handling, how can one evaluate and quantify the trade-off between the three approaches? for the same level of performance, what is the difference in required capacity between the three approaches and how can this be used to justify the choice of one approach over another? 5/46

6 Questions How much more network capacity is needed with aggregate versus per-flow handling? How does the complexity of per-flow handling compare to capacity costs of aggregate traffic handling? How sensitive are the capacity requirements to variations in delay requirements? 6/46

7 Introduction and Motivation Approach Alternatives Simulation Stochastic Analysis Deterministic Analysis Chose deterministic analysis using network calculus to provide bounds on capacity Not dependent on traffic models Suitable for architectural comparisons 7/46

8 Network Calculus Traffic Characterization Arrival curves: upper bound on amount of traffic in a given interval Example: IETF arrival curve model A(t) = min(m + pt; b + rt) M = maximum packet size, p= peak rate, b = burst tolerance, r = sustainable rate 8/46

9 ( Network Calculus Network Characterization Service curves: lower bound on amount of traffic that receives service Example: IETF rate-latency service curve S(t) = R[t T ] + = R(t T ) t T 0 otherwise R is the guaranteed rate and T is the latency 9/46

10 Network Calculus IETF Arrival and Service Curves Bytes Arrival Curve slope r A(t) slope R dmax S(t) b M slope P wmax Service Curve T Time (sec) delay bound d max backlog bound w max 10/46

11 Network Calculus Traffic Handling Schemes Studied Classification Mechanisms Abbreviation Best-Effort First-In-First-Out FIFO Class-Based Strict Priority Queueing PQ Class-Based Queueing CBQ Per-Flow Weighted Fair Queueing WFQ 11/46

12 t 1 = ffλ t 2 t 1 = ffλ FIFO max = ff d C ff Λ ff ρ C Network Calculus Models of Network Elements: FIFO Scheduler A(t) = ff + ρt S(t) = Ct Bits slope Arrival Curve C 2 = ffλ ff t Service Curve * dmax slope C ρ t 1 t 2 Time (sec) 12/46

13 ffλ ff t = 1 ρ ffλ + L t = 2 g Network Calculus Models of Network Elements: WFQ Scheduler Bits A(t) = ff + ρt slope Arrival Curve Service Curve S(t) = R(t (L=g + L max =C)) slope g * dmax L max + C t 1 T = L/g + Lmax/C t 2 Time (sec) WFQ max = ff + L d g L max + C 13/46

14 Methodology Applications Application RT/NRT QoS Avg. Rate Burstiness Packet (Mbps) (Bytes) Size (Bytes) Telephony RT low delay Interactive Video RT low delay NRT delay tolerant WWW NRT delay tolerant and WWW are delay-tolerant BUT may require some guaranteed bandwidth to prevent starvation 14/46

15 Methodology Comparison of Capacity Requirements Use WFQ as reference to set network load Find capacity required by CBQ, PQ and FIFO to support same traffic as WFQ Analysis is deterministic thus delay bounds may be loose BUT we are interested in difference in capacity 15/46

16 Methodology Cases Studied Impact of voice or www load on capacity Impact of growth in traffic beyond design point on delay QoS Projection of required capacity with annual traffic growth based on industry estimates Impact of changing delay QoS bounds on capacity Impact of changing burstiness on capacity Impact of uncertainty in delay parameters on capacity 16/46

17 N k = $ ρ ffk + L k % Single-Link Analysis WFQ Equations guaranteed rate WFQ k g = max Number of sources D k ; ρ k ff k Λ C w WFQ g k 17/46

18 C CBQ = PX KX N k ff k + L p D class p Single-Link Analysis Capacity Calculations WFQ Capacity C WFQ = k g WFQ k N CBQ k=1 X p=1 k2p 18/46

19 PQ = max C p=1:::p 9 N k ρ k =; j class 2 k X C FIFO = KX k=1 N k ff k + L max(p) N k ff k D min D class p Single-Link Analysis Capacity Calculations PQ 8< p X j=1 : k 2 class j X p 1 X + FIFO j=1 19/46

20 Single-Link Results Projections on Traffic Growth Number of OC 3 Links WFQ CBQ PQ Projected Capacity with 15%/Yr increase in Voice and 100%/Yr increase in WWW Network Dominated by Voice Number of OC 3 Links WFQ CBQ PQ Projected Capacity with 15%/Yr increase in Voice and 100%/Yr increase in WWW Network Dominated by WWW Year Year Number of OC 3 Links FIFO Number of OC 3 Links FIFO Year Year WFQ capacity increases by factor of 4-7 CBQ capacity increases by factor of 3-5 FIFO capacity increases by factor of /46

21 Single-Link Results Capacity as a function of Voice Delay Delay C WFQ C CBQ =C WFQ C PQ =C WFQ C FIFO =C WFQ (sec) (Mbps) Voice Increase in voice delay = decrease in capacity linear relationship for FIFO CBQ & PQ capacity comparable to WFQ 21/46

22 Single-Link Results Capacity as a function of burstiness Link Capacity (x OC 3) Link Capacity with Video = 20% Burst lengths = 10msec WFQ CBQ PQ Voice Load = 0.05 Voice Load =0.225 Voice Load = 0.4 Voice Load =0.575 Voice Load = 0.7 Link Capacity (x OC 3) Link Capacity with Video = 20% Burst lengths = 10,50,100,100msec WFQ CBQ PQ Voice Load = 0.05 Voice Load =0.225 Voice Load = 0.4 Voice Load =0.575 Voice Load = 0.7 Link Capacity (Number of OC3 links) Link Capacity (Number of OC3 links) WFQ CBQ PQ FIFO 0 WFQ CBQ PQ FIFO smaller burst size = same order of magnitude capacity increasing and WWW burst size increases FIFO capacity significantly 22/46

23 Edge-Core Network Analysis Edge-Core Topology Topology defined by # of core nodes N core &#of links per core n node link core 2) topologies (N E E E E per N core C C C Full-mesh: n link = E C E E E (N core 1) Fixed # of sources per edge node Edge Node Core Node Edge Link Core Link 23/46

24 Edge-Core Network Analysis Analysis of Capacity Requirements use a network with WFQ in both the edge and core as the reference calculate the amount of traffic that can be supported in a WFQ network compare the capacity required for various combinations of traffic handling schemes in the edge and core 24/46

25 Edge-core Network Results Parameters number of edge nodes per core node N edge = 60=N core, N core = 3::20 routes set-up within the core using Djikstra s shortest path algorithm Traffic within the core was distributed symmetrically maximum load on each edge link w T = 90% 25/46

26 Edge-Core Network Results Network Capacity for 20-node Full-Mesh Core Traffic Handling WFQ CBQ PQ FIFO Edge WFQ Traffic CBQ Handling PQ FIFO Network capacity in equivalent OC-3 links All-FIFO capacity ß 22x all-wfq network All-CBQ capacity ß 2.5x all-wfq network All-PQ capacity ß 1.5x all-wfq network 26/46

27 Edge-Core Network Results Impact of Network Diameter with WFQ Edge Max Hops C WFQ (x OC3) C CBQ =C WFQ C PQ =C WFQ C FIFO =C WFQ 1 (full-mesh) utilization decreases with increasing diameter WFQ: , CBQ: , FIFO: more links => smaller diameter => higher per-node delay => smaller capacity 27/46

28 Edge-Core Network Results Impact of Delay Bound with FIFO Edge Voice Delay Bound C WFQ (Mbps) C CBQ =C WFQ C PQ =C WFQ C FIFO =C WFQ core nodes WFQ capacity increases with increasing voice delay due to increased burstiness in FIFO edge CBQ, PQ and FIFO capacity decreases with increasing voice delay 28/46

29 Simplified Bounds on Capacity Goal simplify capacity calculations for CBQ, PQ and FIFO obtain simple bounds for ratios of CBQ, PQ and FIFO capacity to WFQ capacity as functions of: maximum delay bound aggregate burstiness 29/46

30 C FIFO fl FIFO = fl CBQ = g WFQ + Kk=1 N kff k D min P Kk=1 N k P p L min D Simplified Bounds on Capacity Single Link FIFO C WFQ = flfifo g WFQ CBQ C CBQ P P C WFQ < flcbq P K k=1 N kg WFQ p=1 Kk=1 N kff k D k;p P Kk=1 N k 30/46 P

31 Bounds on Capacity Single Link Results for CBQ Voice Delay Bound g WFQ fl CBQ =g WFQ C CBQ =C WFQ (sec) (Mbps) Simple Exact simple bounds accurate can provide reasonable estimate of CBQ fl CBQ capacity 31/46

32 M (1 (M 1)ff) Capacity, Delay and Utilization Path vs non-path Aggregation with PQ scheduler Path Aggregation: aggregated flows share same end-to-end path e.g. MPLS non-path Aggregation(non-PA) network: non PA E2E = ff + L D C Path Aggregation(PA) network: PA E2E = ff + L D C (1 + ff)m 1 ff equations differ in multiplying factor applied to 32/46

33 Capacity, Delay and Utilization Path vs non-path Aggregation with PQ scheduler Nodes ff = 0:1 Utilization ff = 0:9 Utilization (M) non-pa PA non-pa PA * * non-pa bound large when M = 10 and ff = 0:1 non-pa bound not applicable M > 2 for ff = 0:9 and PA can provide better utilization and delay QoS over large networks 33/46

34 Sensitivity Analysis Motivation for Sensitivity Analysis analysis presented so far requires knowledge of the maximum delay bounds for each traffic type carried in the network must consider how imperfect knowledge affects capacity must consider how traffic handling mechanisms are affected by uncertainty in the delay bounds 34/46

35 Sensitivity Analysis Questions 1. what is the uncertainty in the capacity requirements given the uncertainty in the delay bounds? 2. how important are the individual delay bounds for each traffic type with respect to the uncertainty in the capacity? 35/46

36 EfCg = C KX Sensitivity Analysis Sensitivity Analysis of capacity 2 C V ar[d k ] 2 k k V ar[c] = V ar[d k ] k=1 36/46

37 Sensitivity Analysis Results Parameters D Type (ms) D max (ms) D mean (ms) V ar[d](ms) 2 min Voice Video WWW uniformly-distributed delays overlapping distributions analysis possible for WFQ only 37/46

38 Sensitivity Analysis Results WFQ Results p V Parameter Voice Video WWW Analytic Sensitivity (Mbps/msec) % Variance Simulation Sensitivity (Mbps/msec) ariance(mbps) % Variance sensitivity highest for voice most variance in capacity due to and WWW ranking of % variance same for analytic and simulation 38/46

39 Sensitivity Analysis Results CBQ Simulation Results p V ariance(mbps) Parameter Voice Video WWW Sensitivity (Mbps/msec) % Variance variance in capacity due to and WWW sensitivity highest for voice similar results for PQ 39/46

40 Sensitivity Analysis Results FIFO Simulation Results p V Parameter Voice Video WWW Sensitivity (Mbps/msec) ariance(mbps) % Variance sensitivity highest for voice variance in capacity due to voice and video 40/46

41 Lessons Learned Can quantify comparison of capacity requirements CBQ & PQ network capacity ß WFQ capacity importance of network architecture in comparing traffic handling approaches edge-core traffic handling combinations path vs non-path aggregation sensitivity analysis helps to identify critical parameters affecting capacity requirements quantifying complexity vs cost of capacity non-trivial task 41/46

42 Significance of Results Network Architecture all-wfq = small capacity + high complexity Flow-based (WFQ) Core Traffic Handling Class-based (CBQ/PQ) Best-Effort (FIFO) Combination of WFQ, CBQ, PQ = small capacity + medium to high complexity Edge Traffic Handling Flow-based (WFQ) Class-based (CBQ/PQ) all-fifo = huge capacity + least complexity Best-Effort (FIFO) FIFO + (WFQ,CBQ,PQ) = moderate capacity + medium to high complexity KEY Increasing Network Capacity 42/46

43 Significance of Results Network Design Simplified Bounds on capacity use simpler WFQ analysis to obtain estimates of CBQ, PQ and FIFO capacity Path-aggregation analysis relationship between capacity, utilization and delay can be used to illustrate benefits of architectures such as MPLS that use path-aggregation 43/46

44 Significance of Results Sensitivity Analysis Planning and Forecasting impact of changes in traffic patterns on capacity Service Provisioning types of services possible with existing capacity Network Management choice of parameters for desired service 44/46

45 Summary of Contributions developed methodology for comparing different traffic handling approaches demonstrated equivalence of class-based schemes and flow-based schemes wrt capacity requirements Per-flow traffic management vs Differentiated services debate developed sensitivity analysis for long-term network planning extended the application of Network Calculus in addressing a significant networking problem 45/46

46 Future Work use of stochastically bounded traffic models obtaining better bounds on utilization for networks that use aggregate traffic handling extend work on capacity bounds sensitivity analysis obtaining an analytic solution using order statistics use of global sensitivity analysis such as the importance measures 46/46