SRPT Applied to Bandwidth Sharing Networks (to appear in Annals of Operations Research) Samuli Aalto TKK Helsinki University of Technology, Finland Urtzi Ayesta LAAS-CNRS, France BSnetworks.ppt TKK/ComNet Research Seminar, 7.3.2008 1 Outline • • • • • Introduction: Bandwidth sharing networks Background: Static setting Background: Dynamic setting Theoretical results: Delay reduction by applying SRPT Numerical results and observations 2 Resource sharing • Single resource – loss system (e.g. Erlang model M/G/n/n) • one customer per server, no buffer • model for telephone traffic at the connection level – queueing system (e.g. M/G/1-FCFS) • customers served one-by-one, infinite buffer • model for data traffic at the packet level – sharing system (e.g. M/G/1-PS) • customers share the server, infinite buffer • model for (elastic) data traffic at the flow level • Multiple resources – loss networks (e.g. Ross) – queueing networks (e.g. Jackson, BCMP, Kelly, ...) – bandwidth sharing networks (Massoulié & Roberts) 3 Bandwidth sharing network • Characterization: – Flow-level model of a data network loaded with elastic flows (such as file transfers using TCP) • Resources: – Links l with bandwidth cl (bps) – Routes r (route = collection of consecutive links) • Traffic: – Elastic flows i with arrival time (possibly random), size (bits to be transferred), and route (fixed) • Control: – Bandwidth allocation to flows • inter-route bw allocation • intra-route bw allocation 4 Example: Symmetric linear network with L = 2 route 1 route 2 link 1 link 2 route 0 Symmetric with unit capacities: c1 c2 1 5 Bandwidth allocation • Inter-route bandwidth allocation – fr = total bandwidth allocated to the flows on route r – feasible allocation satisfies the capacity constraints: rR(l ) fr cl for all l • Intra-route bandwidth allocation – tells how the total bandwidth fr is shared among the flows using route r – an example: • PS = Processor-Sharing = bandwidth is shared evenly among the flows with the same route 6 Outline • • • • • Introduction: Bandwidth sharing networks Background: Static setting Background: Dynamic setting Theoretical results: Delay reduction by applying SRPT Numerical results and observations 7 Static setting • Characterization: – Fixed number of saturated flows, n (nr; r R) • Problem: – Fair bandwidth sharing • Solutions (with PS intra-route discipline): – MMF: max-min fairness [a ]; trad., Jaffe (1981) – PF: proportional fairness [a 1]; Kelly (1997) – TM: throughput maximization [a ]; Massoulié & Roberts (1999) – PDM: potential delay minimization [a 2]; Massoulié & Roberts (1999) – alpha-fairness [a ]; Mo & Walrand (1998,2000) – U-utility maximization; Ye et al. (2003,2005) – BF: balanced fairness; Bonald & Proutière (2003) 8 Fairness in symmetric linear network with L = 2 • alpha-fairness [a ] n0 f0 , 1 / a a a n (n n ) 0 1 • TM [a ] 2 f1 f2 ( n1a n2a )1 / a n0 ( n1a n2a )1 / a n1n2 0 0, n f0 n n0 n , n1n2 0, n1 n2 0 1,0 1 2 n n 0 1 2 • PF [a 1] = BF (in this case, not generally) n0 n1 n2 f0 n n n , f1 f2 n n n 0 1 2 0 1 2 • MMF [a ] n0 , max{ n , n } 0 1 2 f0 n max{n1,n2 } 0 max{n1,n2 } f1 f2 n 9 Fairness in symmetric linear network with L = 2 • Note: Throughput maximization does not specify a unique bandwidth allocation when n1 or n2 • TM as a limit a n1 0 and n2 0 0, n0 f0 n n n , either n1 0 or n2 0 1,0 1 2 n 0 and n 0 1 2 • TM* with preemptive priority to routes 1 and 2 0, f0 0, 1, n1 0 and n2 0 either n1 0 or n2 0 n1 0 and n2 0 10 Outline • • • • • Introduction: Bandwidth sharing networks Background: Static setting Background: Dynamic setting Theoretical results: Delay reduction by applying SRPT Numerical results and observations 11 Dynamic setting • Randomly varying number of flows – Poisson arrivals, generally distributed flow sizes • Necessary stability conditions: rR(l ) r cl for all l • Definition: – Bandwidth allocation policy is (maximally) stable if the necessary stability conditions are also sufficient • Primary concern: – stable bandwidth sharing • Secondary concern: – (mean) delay optimization among stable bandwidth allocations 12 Single (bottleneck) link • M/G/1 queue • Fair bandwidth sharing: PS (Processor-Sharing) • Stability: WC (Work-Conserving disciplines) • Anticipating mean delay optimization: SRPT (ShortestRemaining-Processing-Time); Schrage (1968) • Non-anticipating mean delay optimization for DHR (Decreasing Hazard Rate) service times: FB (ForegroudBackground) = LAS (Least-Attained-Service); Yashkov (1987) 13 Stable bw allocations for multilink networks • Exponential flow sizes – PF stable in linear networks; Massoulié & Roberts (1998) – MMF stable in linear networks; De Veciana et al. (1999) – alpha-fair bw allocations stable for any topology; Bonald and Massoulié (2001) – U-utility maximization bw allocations stable for any topology; Ye et al. (2003,2005): • General flow sizes – BF stable for any topology; Bonald and Proutière (2003) – MMF stable for any topology; Bramson (2005) – PF stable for any topology; Massoulié (2005) – alpha-fair bw allocation stable for tree topologies; Gromoll and Williams (2006) • Note: Above, intra-route discipline always PS 14 Stable bw allocations for multilink networks • Verloop et al. (2006): – PR0 and PR12 stable in symm. linear network • PR0 gives preemptive priority to class 0 whenever nonempty • PR12 gives preemptive priority to classes 1 and 2 whenever both of them are nonempty; otherwise preemptive priority is given to class 0 – Intuitive argument: • Both policies ensure that the whole capacity of each link is used whenever there are flows loading the link • Note: PR12 TM* which gives preemptive priority to classes 1 and 2 whenever at least one of them is nonempty 15 Unstable bw allocations for multiple link networks • Bonald and Massoulié (2001): – TM* not maximally stable in linear network – TM* stable in symmetric linear network only if 0 (1 1)(1 2 ) min{1 1,1 2} • Verloop et al. (2005): – global SRPT not maximally stable in linear network – global LAS not maximally stable in linear network • Note: In all these cases, the whole capacity of a link is not necessarily used while there are flows loading the link 16 Delay optimization among stable bw allocations • Yang & de Veciana (2002,2004): – optimal allocation: frn or 1 (depending on state n) in symmetric linear network • Verloop et al. (2006): – determined optimal non-anticipating allocation in symmetric linear network with exponential flow sizes – if m1 m2 m0, then PR0 optimal – if m1m2 m0 and m1 m2 m0, then PR12 optimal • Bonald and Proutière (2003): – BF insensitive for any topology 17 Outline • • • • • Introduction: Bandwidth sharing networks Background: Static setting Background: Dynamic setting Theoretical results: Delay reduction by applying SRPT Numerical results and observations 18 Theoretical setup • Consider a BS network with – a general topology, – Poisson arrivals, and – generally distributed flow sizes • P = family of stable bw allocation policies p 19 Delay reduction by local SRPT’ • Pn = family of stable bw allocation policies p for which Z rp (t ) frp (Np (t )) where – Zr(t) = total bw allocated to class r at time t – Nr(t) = number of flows on route r at time t – N(t) (Nr(t); r R) • Note: All fair bw allocation policies mentioned above Pn 20 Delay reduction by local SRPT’ • Let p Pn be fixed. • p~ = a modified policy – with the same inter-route bw allocation process, p~ Z r (t ) Z rp (t ) frp (Np (t )) – but the intra-route disciplines may be different from the original ones • p’ • p* = the modified policy – that applies SRPT as the intra-route discipline = the policy – for which the inter-route bw allocation process is Z rp * (t ) frp (Np * (t )) – and that applies SRPT as the intra-route discipline 21 Delay reduction by local SRPT’ • Note the difference between p’ and p* • Proposition 1: – Let p Pn, r R and t . ~ For any modification p (including p ), p p~ N r (t ) N r (t ) • Theorem 1: – Let p Pn. ~ For any modification p (including p ), p p~ E[T ] E[T ] • Here T refers to delay (= total transfer time) 22 Delay reduction by local SRPT* • Pw = family of stable bw allocation policies p for which Z rp (t ) rp ( Wp (t )) where – Zr(t) = total bw allocated to class r at time t – Wr(t) = total workload on route r at time t – W(t) (Wr(t); r R) • Note: Policies PR0 and PR12 Pw 23 Delay reduction by local SRPT* • Now there is no difference between p’ and p* • Proposition 2: – Let p Pw, r R and t . ~ For any modification p (including p ), p* p p~ N r (t ) N r (t ) N r (t ) • Theorem 2: – Let p Pw. ~ For any modification p (including p ), E[T p* p p~ ] E[T ] E[T ] 24 Outline • • • • • Introduction: Bandwidth sharing networks Background: Static setting Background: Dynamic setting Theoretical results: Delay reduction by applying SRPT Numerical results and observations 25 Simulation setup • Symmetric linear network with L = 2 and unit capacities – Poisson arrivals with constant total rate l 1 – Flow size distribution with mean b .8 • deterministic • exponential: m 1/b • hyperexponential: p1 .9, m1 9/b; p2 .1, m2 1/9b • Now, PR12 is the optimal non-anticipating allocation policy • Mean delay comparison between basic policy p and its SRPTmodifications p’ and p* using basic policies – BF = PF, PR0, PR12 26 Deterministic flow sizes p p’ p* Mean number of flows 4 3.5 3 BF 2.5 2 1.5 PR12 1 0.5 0.2 0.4 0.6 lambda0 0.8 1 Mean number of flows 4 3.5 3 2.5 2 1.5 PR12 1 0.5 0.2 0.4 0.6 lambda0 0.8 27 1 Exponential flow sizes p p’ p* Mean number of flows 4 3.5 3 BF 2.5 2 1.5 PR12 1 0.5 0.2 0.4 0.6 lambda0 0.8 1 Mean number of flows 4 3.5 3 2.5 2 1.5 PR12 1 0.5 0.2 0.4 0.6 lambda0 0.8 28 1 Hyperexponential flow sizes p p’ p* Mean number of flows 4 3.5 3 BF 2.5 2 1.5 PR12 1 0.5 0.2 0.4 0.6 lambda0 0.8 1 0.4 0.6 lambda0 0.8 1 Mean number of flows 4 3.5 3 BF 2.5 2 1.5 1 0.5 0.2 29 Increasing variability 5 6 BF Mean number of flows Mean number of flows 6 4 3 2 1 0.5 1 1.5 2 2.5 3 3.5 Coefficient of variation 4 5 PR0 4 3 2 1 0.5 1 1.5 2 2.5 3 3.5 Coefficient of variation 4 Mean number of flows 6 5 PR12 4 3 2 1 0.5 1 1.5 2 2.5 3 3.5 Coefficient of variation 4 30 Observations • Mean delay improved by p’ for each class as Prop 1 predicts • As Prop 2 implies, for the basic policies PR0 and PR12 N rp * N rp N rp • In all numerical cases, for the basic policy BF N rp * N rp • In all numerical cases, for all basic policies (BF, PR0, PR12) Np* Np Np • Basic policy PR12 is better than BF for deterministic and exponential flow sizes but worse for hyperexpontial • Delay reduction of BF very similar for exponential and hyperexponential flow sizes 31 THE END 32