Finite Buffer Fluid Networks with Overflows Yoni Nazarathy, Swinburne University of Technology, Melbourne. Stijn Fleuren and Erjen Lefeber, Eindhoven University of Technology, the Netherlands. Talk Outline • Background: Open Jackson networks • Introducing finite buffers and overflows –Interlude: How I got to this problem • Fluid networks as limiting approximations • Traffic equations and their solution • Almost discrete sojourn times Open Jackson Networks Jackson 1957, Goodman & Massey 1984, Chen & Mandelbaum 1991 Problem Data: , , P 1 Assume: open, no “dead” nodes Traffic Equations (Stable Case): M i i pi j M p i 1 pi j j 1 i i j p j i j 1 P ' ( I P ') 1 Traffic Equations (General Case): i i j j p j i M M j 1 P ' LCP ( I P ') , ( I P ') Open Jackson Networks Jackson 1957, Goodman & Massey 1984, Chen & Mandelbaum 1991 Problem Data: , , P 1 Assume: open, no “dead” nodes Traffic Equations (Stable Case): M i i pi j M p i 1 pi j j 1 i i j p j i j 1 P ' ( I P ') 1 Product Form “Miracle”: M j lim P X 1 (t ) k1 ,..., X M (t ) kM 1 t j 1 j M j j kj Modification: Finite Buffers and Overflows Problem Data: K1 , , P, K , Q 1 Assume: open, no “dead” nodes, no “jam” (open overflows) qi j i Ki i pi j Explicit Solutions: M p i 1 pi j j 1 M qi 1 qi j j 1 K M M Generally No Exact Traffic Equations: Generally No A Practical (Important) Model: Yes Our Contribution (in progress) Limiting Traffic Equations: K1 1 qi j i Ki P ' Q '( ) i Efficient Algorithm for Unique Solution: pi j M p i 1 pi j j 1 M qi 1 qi j j 1 K M M Limiting Deterministic Trajectories X ( N ) (t ) lim supt x(t ) 0 N N Limiting Sojourn Time Distribution S (N ) S P(S k ) 1 T k 1 Interlude: How I got to this problem Control of queueing networks: Server 1 Server 2 PUSH PULL 1 1 PULL 2 PUSH 2 Output process, D(t), asymptotic variance: E D(t ) lim t t vs. Var D(t ) lim t t 2 3 BRAVO effect for M/M/1/K load When K is Big, Things are “Simpler” for K big, out rate overflow rate ( ) Scaling Yields a Fluid System A sequence of systems: N 1, 2,... (N ) (N ) N (N ) N K Make the jobs fast and the buffers big by taking N The proposed limiting model is a deterministic fluid system: Fluid Trajectories as an Approximation (N ) X (t ) lim supt x(t ) 0 N N Traffic Equations (at equib. point) out rate overflow rate ( ) i i j j p ji j j q ji M M j 1 j 1 or P ' Q '( ) or LCP ( I Q ') ( ( I P ') ) , ( I Q ') ( I P ') 1 1 LCP (Linear Complementarity Problem) a M ,G M M LCP (a, G ) :Find z , w M such that, w Gz a, w 0, z 0, w ' z 0. The last (complemenatrity) condition reads: wi 0 zi 0 and zi 0 wi 0. Min-Linear Equations as LCP Find : B( ) B 0 0 ( ) '( ) 0 w ( I B) z ( I B) z 0, w 0 w' z 0 w , z LCP( ( I B) , I B) Existence, Uniqueness and Solution Definition: A matrix, G M M is a "P"-matrix if the determinants of all (2n 1) principal submatrices are positive. Theorem (1958): LCP(a, G) has a unique solution for all a M if and only if G is a "P"-matrix. "P"-matrix means that the complementary cones "parition" Immediate naive algorithm with 2M steps 1 0 w1 g11 0 1 w g 2 21 g12 z1 a1 g 22 z2 a2 0 1 We essentially assume that our 1 matrix ( G (I Q ') ( I P ') ) is a “P”-Matrix We have an algorithm (for our type of G) taking M2 steps n C{1,2} C{2} 1 0 g 11 g 21 C g 12 g 22 a 1 a2 C{1} Sojourn Time Time in system of customer arriving to steady state FCFS system S ( N ) Sojourn time of customer in N'th scaled system We want to find the limiting distribution of S (N ) Sojourn Times Scale to a Discrete Distribution!!! PS (N ) x x “Molecule” Sojourn Times F {1,..., s} i i for i F F {s 1,..., M } i i for i F Observe, time through i F For job at entrance of buffer iF w. p. : i time through i F 0 i enters buffer i i w. p. 1 i i w. p. 1 i i A job at entrance of buffer Ki iF: A “fast” chain and “slow” chain… qij routed to entrace of buffer j q i leaves the system routed almost immediately according to P The “Fast” Chain and “Slow” Chain Example: M 4 , 4 1, i 1, 1 2 i 1 K1 K2 F {1, 2}, F {3, 4} 4 a 4 p start j 1 4 a j 1 j 1 1 ji : 1’ 1 1j a j1 j j0 j 11 1 q1 4 1 p1 p1 j a j 0 j 1 4 p “Fast” chain on {0, 1, 2, 1’, 2’, 3’, 4’}: ai j Absorbtion probability in j {0,1, 2} starting in i' j 1 2’ 1 1 q1i 1 1j a j2 0 2 3’ “Slow” chain on {0, 1, 2} DPH distribution (hitting time of 0) transitions based on “Fast” chain p4 i p4 4’ E.g: Moshe Haviv (soon) book: Queues, Section on “Shortcutting states” The DPH Parameters (Details) F {1,..., s}, F {s 1,..., M } “Fast” chain BM s 1 0 1 s 0 s 0M s s CM M 1 1 1 0 M s s 0 0 M M s Q 0 M s s M s P I M s 1 s s AM s (I C)1 B “Slow” chain Tss I s 0 s M s P A s1 1 M j 1 S ~ DPH (Tss ,1s ) T A j P(S k ) 1 T 1s1 k Sojourn Times Scale to a Discrete Distribution!!! “Almost Discrete” Sojourn Time Phenomena Taken from seminar of Avi Mandelbaum, MSOM 2010 (slide 82). Summary – Trend in queueing networks in past 20 years: “When don’t have product-form…. don’t give up: try asymptotics” – Limiting traffic equations and trajectories – Molecule sojourn times (asymptotic) – Discrete!!! – Future work on the limits.