An efficient ans simple class M. Boyer An efficient and simple class of functions to model arrival curve of packetised flows Network calculus Shaping, packetization and computation time Marc Boyer, Jörn Migge, Nicolas Navet Swaping between function classes Experiment Conclusion M. Boyer (ONERA,France) Journée WEED 23 Novembre 2011 An efficient ans simple class WEED - Nov. 2011 1 / 22 Outline An efficient ans simple class M. Boyer 1 Network calculus Network calculus 2 Shaping, packetization and computation time Shaping, packetization and computation time 3 Swaping between function classes Swaping between function classes Experiment 4 Experiment 5 Conclusion Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 2 / 22 Outline An efficient ans simple class M. Boyer 1 Network calculus Network calculus 2 Shaping, packetization and computation time Shaping, packetization and computation time 3 Swaping between function classes Swaping between function classes Experiment 4 Experiment 5 Conclusion Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 3 / 22 Network calculus overview An efficient ans simple class M. Boyer Network calculus Two basic objects: Flow: modelling: R ∈ F = {R+ → R+ , non-decreasing} semantics: R(t), cumulative amount of data up to t Server: S modelling: S ∈ F × F: R − → R 0 =⇒ R 0 ≤ R semantics: relation between some input and some output, no loss, output comes after input (R 0 (t) ≤ R(t)) Shaping, packetization and computation time Swaping between function classes Experiment Conclusion delay: R d(t) b(t) R’ v(R,R’) max h(R, R 0 ) S R− →R 0 0 h(R, R ) : horizontal deviation d(R, S) = h(R,R’) t M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 4 / 22 Contract modelling An efficient ans simple class Flow contract: arrival curve α R ≺ α ⇐⇒ ∀t, ∆ ∈ R+ R(t + ∆) − R(t) ≤ α(∆) M. Boyer ⇐⇒ R ≤ R ∗ α Network calculus Server contract: service curve simple service of curve β S R− → R 0 ⇐⇒ R 0 ≥ R ∗ β strict service of curve β for all backlogged period [t, t + ∆[ (i.e.∀x ∈ [t, t + ∆[: R 0 (x) < R(x)): R 0 (t + ∆) − R 0 (t) ≥ β(∆) Shaping, packetization and computation time Swaping between function classes Experiment S Results: R − → R 0 , R ≺ α, S has service curve β: Conclusion R0 ≺ α β d(R, S) ≤ h(α, β) M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 5 / 22 Outline An efficient ans simple class M. Boyer 1 Network calculus Network calculus 2 Shaping, packetization and computation time Shaping, packetization and computation time 3 Swaping between function classes Swaping between function classes Experiment 4 Experiment 5 Conclusion Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 6 / 22 Shaping on links An efficient ans simple class M. Boyer Network calculus Shaping, packetization and computation time Swaping between function classes A link is shared by a set of flows: what is the throughput of this set ? Principle: whatever the applicative throughput is, is it limited by the links capacity Also known has: Serialisation: the frames of the different flows can not be sent at the same time Grouping: computes per-group throughput, not per-flow Interest: considering long term rate ρ and instantaneous burst b applicative flows: small ρ, big b link: big ρ, null b Experiment Conclusion Impact: up to 40% in industrial system M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 7 / 22 Shaping and network calculus An efficient ans simple class Kb Shaping M. Boyer Group sum Network calculus Shaping, packetization and computation time Swaping between function classes ms Let S be a server, with shaping curve σ, then, the output is constrained by σ. If the output is constrained by α0 , it is by α0 ∧ σ. Experiment Conclusion S R− → R 0 =⇒ R 0 ≺ σ R ≺ α, S D β =⇒ R 0 ≺ σ ∧ (α β) M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 8 / 22 Modelling a packetized flow An efficient ans simple class M. Boyer Network calculus Shaping, packetization and computation time Common example: sporadic flow inter emission “period”: T frame size (fixed on min): b Two modelling: fluid (“token bucket”): affine function, continuous packetized: stairs functions, discontinuous Kb Swaping between function classes Fluid Experiment Frame Conclusion Packet Size ms T M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 9 / 22 Fluid modelling: the virtual burst problem An efficient ans simple class M. Boyer Network calculus Jitter “shifts” the arrival curve: if jitter < period: instantaneous burst unchanged in fluid modelling: creation of virtual burst =⇒ increase bounds Shaping, packetization and computation time Swaping between function classes Experiment Kb Fluid Virtual Burst Frame Packet Size Conclusion ms T − Jitter M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 10 / 22 Putting all together An efficient ans simple class M. Boyer Network calculus Shaping, packetization and computation time Swaping between function classes Experiment Conclusion fluid + shaping: concave piecewise linear function (CPL) Efficient min, max, sum Implementation in floating points stair-case modelling: general class (UPP) Complex min, max, sum Implementation in exact rationals (Q) M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 11 / 22 Outline An efficient ans simple class M. Boyer 1 Network calculus Network calculus 2 Shaping, packetization and computation time Shaping, packetization and computation time 3 Swaping between function classes Swaping between function classes Experiment 4 Experiment 5 Conclusion Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 12 / 22 Getting the better of each class An efficient ans simple class M. Boyer Classes strengths/weaknesses: jitter effect: stair-case class Network calculus Shaping, packetization and computation time Swaping between function classes summing (“grouping”): CPL class shaping: CPL class Idea: keeping stair-case for individual flow constraint converting into CPL when summing and shaping Experiment Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 13 / 22 From stair-case to CPL An efficient ans simple class M. Boyer γ b T −τ ,b γ b ,b(1+τ /T ) νT ,τ Network calculus b Shaping, packetization and computation time Swaping between function classes T T −τ t Figure: CPL overapproximation of a stair-case function Experiment Conclusion M. Boyer (ONERA,France) cpl(bνT ,τ ) = γ b nT −τ ,nb An efficient ans simple class ∧ γ b ,b(1+τ /T ) (1) T WEED - Nov. 2011 14 / 22 Algorithm adaptation An efficient ans simple class M. Boyer Network calculus Shaping, packetization and computation time Swaping between function classes Experiment Conclusion Adaptation: replace M. Boyer (ONERA,France) P Fik ∈F An efficient ans simple class αik by P Fik ∈F cpl(αik ) WEED - Nov. 2011 15 / 22 Outline An efficient ans simple class M. Boyer 1 Network calculus Network calculus 2 Shaping, packetization and computation time Shaping, packetization and computation time 3 Swaping between function classes Swaping between function classes Experiment 4 Experiment 5 Conclusion Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 16 / 22 Testbed configuration An efficient ans simple class M. Boyer Network calculus Shaping, packetization and computation time Swaping between function classes industrial (Thales) configuration 104 nodes 8 switches 974 multicast flows 6501 end-to-end bounds Experiment Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 17 / 22 Comparing methods An efficient ans simple class 12000 M. Boyer 10000 Network calculus Shaping, packetization and computation time Swaping between function classes 8000 CPL CPL/nu UPP 6000 4000 2000 Experiment Conclusion 0 326 652 M. Boyer (ONERA,France) 978 1304 1630 1956 2282 2608 2934 3260 3586 An efficient ans simple class 3912 4238 4564 4890 5216 5542 5868 6194 WEED - Nov. 2011 18 / 22 Zoom on worst delays An efficient ans simple class 11000 M. Boyer 10000 Network calculus Shaping, packetization and computation time 9000 CPL CPL/nu UPP 8000 Swaping between function classes 7000 Experiment Conclusion 6000 51 102 M. Boyer (ONERA,France) 153 204 255 306 357 408 459 510 561 An efficient ans simple class 612 663 714 765 816 867 918 969 WEED - Nov. 2011 19 / 22 Experiment values An efficient ans simple class Method M. Boyer Network calculus Shaping, packetization and computation time Swaping between function classes Computation time Min gain Max gain Av. gain Min gain on 1000 biggest Max gain on 1000 biggest Av. gain on 1000 biggest CPL (float) 0.9 s - CPL/b.νT ,τ (float) 1.1 s 0% 7.8% 2.49% 0.8% 4.4% 2.9% UPP (rat) 7.2 s 0.15% 15.2% 5.92% 2.0% 11.9% 8.3% Experiment Conclusion M. Boyer (ONERA,France) Gain correlation: 0.785 An efficient ans simple class WEED - Nov. 2011 20 / 22 Outline An efficient ans simple class M. Boyer 1 Network calculus Network calculus 2 Shaping, packetization and computation time Shaping, packetization and computation time 3 Swaping between function classes Swaping between function classes Experiment 4 Experiment 5 Conclusion Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 21 / 22 Conclusion An efficient ans simple class M. Boyer Network calculus Shaping, packetization and computation time Swaping between function classes Two critical aspects: shaping and packetization Two existing methods: fluid: bad packetization, quick computation stair-case: good packetization, longer computation Contribution: trade-off tightness/computation time Perspective: use in optimisation loop quick computation in first iterations longer computation to finalise Experiment Conclusion M. Boyer (ONERA,France) An efficient ans simple class WEED - Nov. 2011 22 / 22