SDN: Systemc Rsks due to Dynamc Load Balancng Vladmr Marbukh IRTF SDN
Abstract SDN acltates dynamc load balancng Systemc benets o dynamc load balancng: - economc: hgher resource utlzaton, hgher revenue,.. - reslence/robustness to alures, demand varablty,.. Systemc rsks o dynamc load balancng: - robust to small yet ragle to large-scale alures/overload - possblty o abrupt cascadng overload - persstent/metastable systemcally congested states Necessty to manage SDN systemc rsk/benet tradeo
Congeston-aware Routng n a Delay Network P. Echenque, J. Gomez-Gardenes, and Y. Moreno, Dynamcs o ammng transtons n complex networks, 004. Network congeston level A 4 A 3 A1 A Exogenous load h=1: congeston oblvous (mnmum hop count) routng h=0: congeston aware routng Mnmum-cost routng Route cost: C = hd + ( 1 h) q d q # hops rom node to the destnaton queue length at node Congeston-aware routng robust to small yet ragle to large-scale congeston Benet: lower network congeston or medum exogenous load rom A1 to A Rsk: hard/severe network overload (dscontnuous phase transton) at A Economcs drves system to the stablty boundary A. 3
Congeston-aware Routng n Loss Network Arrvng request s routed drectly possble, otherwse an avalable -lnk transt route. Perormance: request loss rate L. Postve eedback: load ncrease à more transt routes à load ncrease.. = Cascadng overload Fully connected network Combnaton o selsh requests & varable demand => emergence o congested metastable (persstent) state => robust (to local) yet ragle (to large-scale congeston) Loss under mean-eld approxmaton [F. Kelly] Metastablty/Cascadng overload [F. Kelly] 4
Λ 1 B1 c1 Cloud wth Dynamc Load Balancng N 1 where utlzaton s > 0, Λ J ( N c Problems: exogenous load uncertan, other uncertantes. Possble soluton: dynamc load balancng based on dynamc utlzaton, e.g., numbers o occuped servers, queue szes, etc. Problem: servng non-natve requests s less ecent: BJ cj N J = Λ Server group : operatonal wth prob. non-operatonal wth prob. Statc load balancng s possble : ) = 1 O( N c < c, and accordng to A.L. Stolyar and E. Yudovna (013) ths may cause nstablty o natural dynamc load balancng 1 + α 1 Falures/recoveres on much slower tme scale than ob arrvals/departures = 0, and α 0, N ) 5
1 1 0 0 L ~ Dynamc Load Balancng n Cloud [V. Marbukh, 014] B A A A A =1 =θ B B B E 1 E α E 0 Fgure. Lost revenue vs. exogenous load or derent levels o resource sharng L ~ A 0 B A α B A E opt α α 1 Fgure. Provder perspectve: lost revenue vs. resource sharng level. α (a) As level o resource sharng exceeds certan threshold, metastable/persstent congested equlbrum emerges, makng Cloud robust to local overload yet ragle to large-scale overload (b) Wth urther ncrease n resource sharng, perormance o the normal metastable equlbrum mproves, whle o the congested metastable equlbrum worsens. (a) Economcs o the normal equlbrum drves Cloud rom robust to ragle and eventually to stablty boundary o the normal equlbrum. (b) Ths creates nherent tradeo between lost revenue: ~ ~ SysLoss α) = L ( α) L ( α ( and systemc rsk o large scale overload ~ ~ SysRsk ( α) = [ L ( α) L ( α)] P( α opt ) α) 6
Systemc Perormance/Rsk Tradeo n Cloud SysLoss 0 Fgure. Rsk/Perormance tradeo: 1 SysRsk 1 > P( α ) ~ exp[ ( ~ < α α α) SysLoss ~ ( α) ( α α) SysRsk ~ exp 1 SysLoss σ (σ )] Implcaton: Uncertanty makes systemc Rsk/Perormance tradeo essental Queston: How can one-dmensonal analyss descrbe a heterogeneous Cloud? Answer: Perron-Frobenus theory due to congeston dynamcs beng non-negatve Snce normal equlbrum loses stablty as Perron-Frobenus egenvalue o the lnearzed system crosses pont γ =1 rom below, t s natural to quanty the system stablty margn and rsk o cascadng overload by Δ = 1 γ Word o cauton: the above results are obtan under mean-eld approxmaton. 7
Future Research Vercaton/valdaton results obtaned under mean-eld approxmaton through smulatons, measurements on networks and rgorous analyss (doubtul). Possblty o onlne measurement o the Perron-Frobenus egenvalue or the purpose o usng t as a bass or early warnng system. Possblty o controllng networks, especally through prcng, based on the Perron-Frobenus egenvalue. 8
Thank you! 9