Chapter 4 Networks, Series, and Cyclic Queues CONTENTS 4.1 SERIES QUEUES ................................................................................................ - 5 4.1.1 QUEUE OUTPUT ................................................................................... - 6 - 4.1.2 SERIES QUEUES WITH BLOCKING ..................................................... - 12 - 4.2 OPEN JACKSON NETWORKS ........................................................................... - 15 4.2.1 OPEN JACKSON NETWORKS WITH MULTIPLE CUSTOMER CLASSES - 23 - 4.3 CLOSED JACKSON NETWORKS ...................................................................... - 24 4.3.1 MEAN-VALUE ANALYSIS .................................................................... - 29 - 4.4 CYCLIC QUEUES ............................................................................................. - 41 4.5 EXTENSIONS OF JACKSON NETWORKS .......................................................... - 43 4.6 NON-JACKSON NETWORKS ............................................................................ - 44 - The networks we considered are with the following characteristics: r10 node 1 (i) rk0 rij r12 m1 g1 r20 … 2 k m2 i mk g2 j rji gk Arrival process from “the outside” to node i follows a Poisson process with mean rate g i , 1 i k . (ii) The service time distribution at node i is independently and exponentially distributed with parameter mi . (iii) The routing probability that a customer who has completed service at node i will go to next node j is rij , where i 1, 2, ,k , j 0, 1, 2, ,k and ri 0 indicates the probability that a customer will depart from the system from node i. This kind of network is called Jackson networks. -2- Closed Jackson Network (i) g i 0 ; No customer may enter the system from the outside. (ii) rio 0 ; No customer may leave the system. Example: M/M/c/c//K The machine-repair problem 1 1 2 … k i = 1, 2 j = 0, 1, 2 m1 = l m2 = m c-1 Operation condition: Holding time = le-lt 2 gi = 0 r12 = 1 r21 = 1 rij = 0, elsewhere c Repair condition: Service time = me-mt Open Jackson network l , (i) ri 0, i 1 1, (ii) rij 1, 0, 1 i k 1, j i 1 ik, j0 elsewhere elsewhere -3- We also call it as Series Queues We restrict ourselves mainly to Markovian systems: Poisson input Exponential service time rij routing probabilities are state independent -4- 4.1 SERIES QUEUES Station 1 Station 2 1 1 … c1 c2 M/M/c1/∞ Example: (i) Computer networks (ii) Time-sharing system (iii) Registration process (iv) Call control process (v) Clinic physical examination -5- 1 … … … l Station n cn 4.1.1 QUEUE OUTPUT For the first station, its departure process is Poisson with mean rate l , or say, its inter-departure time distribution is exponential with mean 1 l Proof If we observe at the output of the first station in an infinitesimal time interval, its probability distribution is Pr Exact one departure in t p1mt p2 2 mt c 1 n 1 n c pc 1 c 1 mt pncmt n c pn nmt pncmt l n l n ( ) ( ) c 1 m m mt n c n c p0 n ! c !c n c n 1 l n 1 l n 1 ( ) ( ) l c 1 m m mt p0 m n 1 (n 1)! n c c!c ( n 1) c l n l n ( ) ( ) c 1 m l m mt p0 lt n c m n ! c ! c n c n 0 -6- Pr more than one departure in t O t since it contains t which is negligible when it compares 2 with t . Pr no departure in t 1 lt O t This implies that it is, which is at all affected by the exponential service mechanism, a Poisson process with mean rate l (We have proven it before) Hence, all stations are M/M/ci / models. For this cascade model, the average system size is just the summation of the system size of each station, and so is the system time. An alternative approach for the proof T t Consider M/M/c/ queue: N(t) Define N(t) as the number of customers in the system at a time t just after the last departure. Pr N (t ) n pn (Continuous time Markov chain) Define T as the random variable of “time between two successive departures”. -7- Fn (t ) Pr N (t ) n, T t (4.2) C (t ) CDF of T Pr T t (4.3) Clearly, C (t ) 1 Pr T t = 1 Fn (t ) n0 Find Fn (t ) Fn (t t ) (1 lt )(1 cmt ) Fn (t ) lt (1 cmt ) Fn 1 (t ) O(t ), ,n c Fn (t t ) (1 lt )(1 nmt ) Fn (t ) lt (1 nmt ) Fn 1 (t ) O(t ), ,1 n c F0 (t t ) (1 lt ) F0 (t ) O(t ) with boundary condition Fn (0) Pr N (0) n, T 0 pn dFn (t ) (l cm ) Fn (t ) l Fn 1 (t ) dt ,n c dFn (t ) (l nm ) Fn (t ) l Fn 1 (t ) dt ,1 n c dF0 (t ) l F0 (t ) dt -8- (4.4) F0 (t ) p0e lt F1 t l m F1 (t ) l F0 (t ) (l m ) F1 (t ) l p0e l t F1 (t ) Ce l m t F1 (t ) p1e l t l p0e l t , C 0 C (t ) 1 p e l t n m n0 1 e lt C t c t lelt Fn t pn e l t Example M/M/c Checkout Counter M/M/∞ 1 Shopping l Exponential m = 4/3 … l= 40/hr 2 c me-mt 1 = 4 min = 1/15 hr m for each counter (i) What is the minimum number of checkout counters? P the checkout counters Cm 3 -9- l 40 m 15 2.67 (ii) If it is described to add one more, then how is the system performance? It is an M/M/4/ model for the checkout station 1 l ( )4 m 4 1 1 l n 4 p0 ( ) 0.06 l 4! 4 n 0 n! m m Wq 0.019 hr 1.14 min Lq lWq 0.76 L 0.76 l m 0.76 40 For shopping station, L l m 15 3.44 40 43 30 Thus the total number of customers in the system is 30 3.44 33.44 - 10 - If the input to station i comes from external system and from another station, and if each of them are independent and Poisson, then we can say the input is still Poisson with mean rate N li li q ji l j j 1 where li is the mean rate of input process coming from external, and q ji is the probability that the output of station j will become the input of station i. We can use the results of M/M/c/ without question. If the problems are that there are limits on the capacity at a station, we treat some of these types of models. - 11 - 4.1.2 SERIES QUEUES WITH BLOCKING Type 1: Two-station single-server-at-each-station sequential model, no queue is allowed. Note that: the server has one room! l m 1e (i) m1t m 2e m2t If there is a customer completed service at station 1, but at this time station 2 is busy, the customer has to wait until the station 2 is free. (ii) Arrivals at station 1 are turned away if station 1 is busy. Define (n1, n2) as system state, and (n1, n2) Description (0, 0) System is busy only (1, 0) Station 1 is busy only (0, 1) Station 2 is busy only (1, 1) Both station 1 and 2 are busy (b, 1) Station 2 is busy, and customer is finished at station 1 but waiting for station 2 to become available - 12 - Then the transition rate diagram is given by (0,0) m2 m1 (0,1) m2 (b,1) l l (1,1) (1,0) m2 m1 And the state-transition equation at steady state is l p0,0 m 2 p0,1 0 m1 p1,0 l p0,0 m 2 p1,1 0 l m p m p m p 0 2 0,1 1 1,0 2 b ,1 m1 m 2 p1,1 l p0,1 0 m 2 pb ,1 m1 p1,1 0 pn1 , n2 1 - 13 - (4.7) If we let m1 m2 m , then l (l 2 m ) l p p , p p0,0 0,0 0,1 1,0 2 m 2m l2 l2 p0,0 , pb ,1 p0,0 p1,1 2 2 2m 2m 2m 2 p 0,0 3l 2 4 ml 2 m 2 Blocking Probability: PB p1,0 p1,1 pb ,1 W system time for accepted customer E service time accepted 2 5 p0,0 ( ) p0,1 ( ) m 2m p0,0 p0,1 Note : 2 m P(T1 T2 ) ( 1 m 1 m 1 m ) P(T1 T2 ) 21 31 m2 m2 5 2m Type 2: Two stations, single server at each station, sequential model, one customer is allowed to wait between the stations. - 14 - 4.2 OPEN JACKSON NETWORKS g1 node r10 1 g2 r12 2 gi … r20 gj rji … i rij ri0 gk … j rj0 k rk0 g i : The mean arrival rate of the external arrival process to node i, 1 i k ; it is in Poisson distribution. mi : The mean service rate of the exponential service time distribution at node i. rij : Independent of system states, 1 i k 1 , 1 j k , (including back to itself) ri 0 : The probability that a customer will leave the network from node i. There is no limit on queue capacity at any node. Markovian system System state Steady-state State Transition equations The method of stochastic balance is used to write down the equations. System state n1 , n2 , , ni , , nj, , nk is somewhat cumbersome. Simplified system state as shown in Table 4.2, p. 174 - 15 - Flow out Flow in k g k g i i 1 i 1 i n ; i- n n k k k m 1 r i 1 i m r k m r i 1 j 1 n ; i+ j- i 1 i ij n ; i+ i j ii i i0 Note 或兩邊cancel 不算在內 n Thus we obtain the state equation k g i 1 k i k k pn ;i mi rij pn ; i j mi rio pn ; i i 1 j 1 i j i 1 (4.9) k k i 1 i 1 mi (1 rii ) pn g i pn Jackson showed that the solution to these equations is, amazingly, of what has come to be generally called Product form. pn C 1n1 2n2 kn k For each node, k li g i rji l j or λ γ λR , j 1 - 16 - i li mi (4.10) And the solution is pn pn1 , n2 ,..., nk (1 1 ) n1 1 (1 2 ) 2 (1 k ) k n2 (4.11) nk That is, the network behaves as if its nodes are independent M/M/1’s. But, as the matter of fact, Disney (1981) showed that the actual internal flow in these kinds of networks which have feedback is not Poisson. (4.11) is a solution of (4.9) Proof pn C 1n1 2n2 C Notice that Kn k K g i 1 n i net flow-in of the network k l r i 1 i io net flow-out of the network 代入 (4.9), we have k k k gi i n n C C mi rij C mi rio i i 1 j 1 i 1 i 1 i j k n i j ? C k n k m (1 r ) C g n i 1 i ii i 1 - 17 - i li m j k g i mi k k li ? k mi rij mi rio ( mi mi rii g i ) l j mi i 1 mi i 1 i 1 li j 1 i 1 k i j From (4.10) k k i 1 i 1 i j l j g j rij li g j rij li rjj l j k r l ij i i 1 i j l j g j rjj l j 代入 式 k g m m i i i l g r l l r m m r g i i ii i i i ii i i i 0 l l i 1 i 1 i i k We can obtain k ? k l r g i 1 i io i 1 i Thus the proof is accomplished! And C nk 0 k n n 1 2 1 n1 0 (1 i ), i 1 1 2 1 KnK 1 1 i 1, - 18 - i 1, 2, 1 1 k ,k 1 Since Li i L , and Wi i 1 i li The total wait within the network of any customer before its final L departure W g i i (Little's formula for the entire network) i i The above results for Jackson networks can be generalized easily to c-channel nodes ci : the number of servers at node i which is with exponential service time mi . pn pn1 , n2 ,..., nk pn1 pn2 k pnk pni i 1 For the case of ci-channel nodes, the above results for Jackson networks can be generalized easily as ri ni pn p0i a ( n ) i 1 i i k l , ri i mi Ref. (2.24) in p.69 - 19 - (4.12) ni !, ai (ni ) n c ci i i ci !, where ni ci ni ci (4.13) p0i p0i ri ni 1 is such that a ( n ) ni 0 i i Thus, again, what we have is a network that just acts as if each node is an independent M/M/c . In fact, it can be shown that (see Disney 1981) as long as there is any kind of feedback, the internal flows are not Poisson. pn pn1n2 nk pn1 pn2 pnk - 20 - Waiting time distribution For arborescent (tree-like (藤狀的)) flow, no direct or indirect feedback, since the input process of each node is Poisson, the actual waiting time distribution at each node should be the same as that for the M/M/c model. However, for a general Jackson network, the input to each node is not necessarily truly Poisson, so that we can not be sure that qn pn (where qn is the probability of n customers in the system at the arrival point). We can have the virtual waiting time distribution at each node in (2.39). (但是真正的 waiting time distribution 則不得而知,除非是 Arborescent network!) - 21 - Example : 1-p l 1 p c1 = 1 2 c2 ≥ 2 1 3 c3 = 1 Bypass 或 multiple servers 的緣故,故 T1( n ) and T3( n ) are dependent. 其 departure process 是否仍為 Poisson? Melamed (1979) showed that “for single-server Jackson networks with an irreducible routing probability matrix rij , and i 1 (every entering customer eventually leaves the network), the departure process for nodes are Poisson and that the collection over all nodes that yield these Poisson departure processes are mutually independent. For Jackson networks, As long as there is no feedback, flows between nodes and to the outside are truly Poisson. If there is feedback, the output process is not Poisson. But, amazingly, Jackson’s solution still holds. - 22 - 4.2.1 OPEN JACKSON NETWORKS WITH MULTIPLE CUSTOMER CLASSES R : The routing probability matrix for a customer of type-t t t 1, 2, , n λ t λ (t ) , λ (t ) (l1(t ) , l2(t ) , lk(t ) ) ( λ ( t ) : the arrival rate of type-t customers) Lt : The average length for node i with M/M/c model Li (t ) li (t ) n l k 1 - 23 - i (k ) Li 4.3 CLOSED JACKSON NETWORKS If g i and ri 0 are zero for all i (N items travel inside the network), we have a closed Jackson network. From (4.9) which comes from the open network model will become k k k mr p j 1 i 1 (i j ) n ; i j i ij mi 1 rii pn (4.14) i 1 where n (n1, n2 , , nk ) It will certainly not be surprising that the solution to it is still of the form pn C 1n1 2n2 kn Cn (4.15) k Since input rate = output rate for each node, that is k li mi i m j j rji j 1 , i 1, 2, ,k lj We can prove (4.15) is the solution of (4.14) (用課本內容說明) and n N pn (4.16) n1 n2 nk N - 24 - C 1n1 kn 1 k Equivalently, C 1n1 n1 n2 nk N 1 knk CN Ref. to (94a) N k 1 There are possible ways that the N customers can be N distributed among the k node. We denote it by C N , and define G N C ( N ) 1 Thus pn 1 1n1 2n2 G( N ) kn k where GN n1 n2 1n 2n 1 nk N 2 kn k The closed network can be extended to Ci servers at node i. The solution now becomes 1 pn G( N ) in k a (n ) i i 1 i - 25 - i where (ni Ci ) (ni Ci ) ni ! ai (ni ) ni Ci Ci ! Ci G( N ) n1 n2 in k a (n ) nk N i 1 i i i Example: N 2 1 1 a 1 ( n1 ) 2 1- r12 r12 1 r23 2 a2 (n2 ) 1 3 a3 (n3 ) 1 1- r23 Holding time l m2 m3 C1 = 2 C2 = 1 C3 = 1 Thus pn pn1 n2 n3 1 1n1 2n2 3n3 G (2) a1 (n1 ) - 26 - , n1 0 , 1 , n1 2 Find 1 , 2 , 3 and G (2) Output rate = input rate for each node Node 1 : l1 m2 2 (1 r23 ) m3 3 Routing Matrix Node 2 : m2 2 l1r12 # : Node 3 : m3 3 l1 (1 r12 ) m2 2 r23 G(2) 12 2 22 32 12 2 3 13 N K 1 2 3 1 4 4 3 6 N 2 2 2 (2 0 0) (1 1 0) (i) (0 2 0) (1 0 1) (0 0 2) (0 1 1) # are redundant. (ii) From , we seem to be able to conclude that pn is independent of the first selection of 1 , or 2 , …, or k . Thus, we arbitrarily set 2 1. Then 2 1 , 1 m2 , r12l n1 pn1 , n2 , n3 3 m2 (1 r12 r12r23 ) r12 m3 1 m2 1 m2 (1 r12 r12 r23 ) G ( N ) r12l a1 (n1 ) r12 m3 - 27 - n3 An efficient algorithm to find G N - by Buzen (1973) Let f i (ni ) G( N ) in i ai (ni ) , then n1 n2 nK N in k a (n ) i 1 i i k i n1 n2 nk N f (n ) i 1 i i Buzen setup an auxiliary function: g m ( n) m n1 n2 f (n ) nm n i 1 i i m nodes n customers If m = k, n = N, then gk ( N ) G( N ) . We now set up a recursive scheme for calculating gm (n) Suppose we fix nm g m ( n) 0 n1 n2 nm 1 f i (ni ) i 1 m n n fm ( ) n n 0 n n m 1 1 2 n n f m ( ) g m 1 (n ) f i (ni ) i 1 m 1 , n 1, m 2 0 Note that g1 (n) f1 (n) and gm (0) 1. Then we can have gm (n) , and thus G N . - 28 - g0 n 0 ,if n 0 g0 n 1 ,if n 0 Further, this scheme aids in calculating marginal distributions as well. pi n Pr N i n pn1 , Si N n fi n , nk k 1 f j nj G N Si N n j 1 f n GN k Si N n j 1 j i where Si n1 j j ni 1 ni 1 , n 0, 1, , N nk 4.3.1 MEAN-VALUE ANALYSIS Two basic principles: (Closed Jackson queueing network) qn ( N ) pn ( N 1) ,0 n N 1 Little’s formula is applicable throughout the network. For M/M/1: system time W ( L 1) m (the system size + itself) service time For M/M/c : System time W = no adjustment which is the same as above. - 29 - ∴ For the closed network (single server) The first principle: Wi ( N ) Li ( N 1) 1 mi (4.23) where Wi ( N ) : mean system time at node i for a network containing N customers. Li ( N 1) : mean number of customers at node i in a network with (N-1) customers. mi : mean service rate. The second principle: Li ( N ) li ( N ) Wi ( N ) (4.24) Considering (4.23) and (4.24), we can calculate Li ( N ) and Wi ( N ) recursively if we can find li ( N ) . Starting with Li (0) 0 , Wi (1) 1 mi Li ( N ), Wi ( N ) , N 1 Compute li ( N ) : li ( N ) N , Di N 1 rate per customer N Di N - 30 - where Di N : the average delay per customer between successive visit to node i for a network with N customers. To get Di N : vi mi i (instead of using li) k vi v j rji ,1 i k (4.25) j 1 One of the k equations is redundant; we set one of vi (say v ) equal to 1 and obtain vi , 1 i k . vi is the relative throughput through node i. vi mi i , assume v is the one we normalize. Then k D ( N ) viWi ( N ) i 1 , v 1次,其他 node 為 vi 次, 通過 node k 故 D ( N ) viWi ( N ) i 1 - 31 - The MVA algorithm for finding Li ( N ) and Wi ( N ) in a k-node, single-server-per-node network with probability R rij is as follows (i) Solve the traffic equation (4.25) k vi v j rji ,1 i k j 1 Setting v 1 (ii) Initialize Li 0 0 , 1 i N (iii) For n 1, to N, calculate (a) Wi (n) (b) l ( n) 1 Li (n 1) mi n k vW (n) i 1 i , i 1, ,k (assume v 1) i (c) li (n) l (n)vi (d) Li (n) li (n) Wi (n) - 32 - ,i 1, ,k routing Example 4.5 0 2 (v1 , v2 , v3 ) (v1 , v2 , v3 ) 3 1 1 4 1 3 0 3 4 0 0 with N 1 Also check with normalizing-constant method. Solution 2 v1 v2 v3 , 3 3 v2 v1 , 4 1 1 v3 v1 v2 4 3 4 2 Set v2 1, then we have v1 , v3 3 3 Applying step (iii) of the algorithm, we have (a) W1 1 W2 1 1 L1 0 l 1 L2 0 W3 1 m2 - 33 - 1 2 1 1 1 1 L3 0 m3 1 3 (b) l2 1 1 3 vW i i 1 i 1 1 4 1 2 1 1 1 3 2 3 3 (c) l1 1 l2 1 v1 9 4 12 17 3 17 l3 1 l2 1 v3 9 2 6 17 3 17 (d) L1 1 l1 1W1 1 12 1 6 17 2 17 L2 1 l2 1W2 1 9 9 1 17 17 L3 1 l3 1W3 1 6 1 2 17 3 17 - 34 - 9 17 Check with Normalizing Constant Method k mi i m j j rji (4.16) j 1 Therefore we have 2 1 2 2 33 3 3 4 2 2 1 33 2 1 , pn1n2 n3 1 2 , 3 1 1 2 1 2 4 3 3 1 1n1 2n2 3n3 GN 2 9 , n1 n2 n3 N 1 2 2 17 where G 1 1n1 2n2 3n3 1 9 9 n1 n2 n3 1 3 p100 2 2 6 2 1 9 , p001 9 3 , p010 17 17 17 17 17 17 9 9 9 Then we have L1 1 p100 9 2 6 , L2 1 p010 , L3 1 p001 17 17 17 - 35 - MVA algorithm to get marginal steady-state probability for extending to the multiple-server case. Detailed balance equation for M/M/1 l pn 1 m pn pn l pn 1 m Applying it to MVA pi (n, N ) li ( N ) pi (n 1, N 1) mi (4.26) We’ll prove it later where pi (n, N ) is the marginal probability of n in an N-customer system at node i, pi (0,0) 1 For our example, p1 (1, 1) l1 (1) l (1) 12 17 p1 (0, 0) 1 6 17 l l 2 p2 (1, 1) l2 (1) 9 17 p2 (0, 0) 9 17 m2 1 p3 (1, 1) l3 (1) 6 17 p3 (0, 0) 2 17 m3 3 ∴ p1 (0, 1) 11 17 , p2 (0, 1) 8 17 , p3 (0, 1) 15 17 - 36 - Checking with our normalizing-constant solution p1 (0, 1) p010 p001 p1 (1, 1) p100 p2 (1, 1) p010 p3 (1, 1) p001 p2 (0, 1) p001 p100 p3 (0, 1) p100 p010 We can see the results from these two approaches are the same in p.193. The MVA algorithm for multiple-server cases Wi (n) 1 mi n -1 1 ci mi ( j c 1) p ( j, n 1) i j ci i ci 2 1 1 Li (n 1) (ci 1 j ) pi ( j , n 1) ci mi j 0 Thus we have to further know pi j, n 1 , j 0, pi ( j, n) li (n) pi ( j 1, n 1) i ( j ) mi where j i ( j) ci - 37 - , j ci , j ci , ci 2 . So the MVA algorithm for multiple servers: k (i) Solve vi rji v j with setting v 1 j 1 (ii) Initialize Li (0) 0 ; Pi (0,0) 1; Pi ( j,0) 0 , j 1. (iii) For n 1, to N, calculate ci 2 1 (a) Wi (n) 1 L ( n 1) ( c 1 j ) p ( j , n 1) , i i i ci mi j 0 i 1, ,k (b) l (n) n k v W ( n) i 1 i i (c) li (n) l (n)vi (d) Li (n) li (n)Wi (n) (e) pi ( j, n) , i 1, li (n) pi ( j 1, n 1) i ( j ) mi ,k , j 1, - 38 - ,n pi (n, N ) Prove li ( N ) pi (n 1, N 1) mi --- the (4.26) for single-server case. Proof pi (ni ; N ) Pr N i ni | N customers in network n1 n2 ni ni 1 1 1n1 nK N n j G ( N ) in Kn i K Define a complementary marginal cumulative probability as Pi (ni ; N ) Pr N i ni | N in network N j ni n1 n2 i j j ni in i G( N ) in i G( N ) ni ni 1 1 1n1 nK N n j G ( N ) 1 g k 1 ( N j ) G( N ) 0 i g k 1 ( N ni ) g k ( N ni ) Now - 39 - in G ( N ni ) i G( N ) i j kn k pi (ni ; N ) Pi (ni ; N ) Pi (ni 1; N ) in i G( N ) G( N ni ) iG( N ni 1) Thus pi (ni ; N ) ini G ( N 1) pi (ni 1; N 1) G ( N ) ini 1 ∴ G ( N ni ) iG ( N ni 1) G ( N 1 ni 1) iG ( N 1 ni 1 1) iG ( N 1) G( N ) pi (ni ; N ) iG ( N 1) G( N ) pi (ni 1 ; N 1) Since li ( N ) throughput at mode i Pr Server is busy at node i mi P (1 ; N ) mi iG ( N 1) G( N ) i G( N ) G ( N 1) mi pi ni ; N li ( N ) mi pi ni 1 ; N 1 pi (ni ; N ) li ( N ) pi (ni 1 ; N 1) mi - 40 - 4.4 CYCLIC QUEUES 1 n1 m 1e … 2 n2 m1t m 2e k k-1 nk-1 m2t m k 1e nk m k 1 t mke mkt And there are N customers in the system It is a special case of a closed Jackson network, and is called cyclic queue, where ,if j i 1, 1 i k 1 1 rij 1 0 ,if i k , j 1 , elsewhere Thus we have pn1 , n2, , nk 1 1n1 2n2 G( N ) kn k where G( N ) n1 n2 nk N 1n 2n 1 2 kn k but mi 1 i 1 , i 2, 3, mi i ,i 1 mk k - 41 - ,k That is m1 2 1 m2 3 m2 m 2 1 1 m3 m3 Due to redundancy Select 1 m k 1 1 mk Then pn1 , n2, , nk m1N n1 1 n2 n3 G ( N ) m 2 m3 mknk where G( N ) n1 nk N 2n kn 2 k For multiple-server case, we can similarly treat it and obtain the result. - 42 - 4.5 EXTENSIONS OF JACKSON NETWORKS Jackson (1963) 外來的 (i) Open network with state-dependent exogenous arrival processes and state-dependent internal service. The solution is also of product form. But the normalizing constant is difficult to obtain. (ii) Jackson networks with travel time (walk time) between nodes of network, (iii) Multiclass Jackson network in which customers have different classes. ※ Baskett et. al. (1975) obtained the product-form solutions. : : : - 43 - 4.6 NON-JACKSON NETWORKS The only non-Jackson networks we considered here are those where the {rij } are allowed to be state dependent. Example Consider a closed network with three nodes and two customers, and the holding times at each node and exogenous inputs if any to them still remain Exponential and Poisson. Suppose r31 r13 = 1 - r12 1 r12 2 r23 r32 = 1 - r31 r21 = 1 - r23 (i) 3 The output customer from each node will visit the next node which has the fewest customers. (ii) If there is a tie, the customer chooses among the remaining nodes with equal probability. - 44 - m3 m1 r12=0.5 r13=0.5 m2 200 020 0.5m1 0.5m1 002 110 101 m3 0.5m2 0.5m2 m1 011 m2 0.5m3 0.5m3 0 m1 p200 0 m 2 p020 0 m3 p002 0 ( m 2 m3 ) p011 1 1 m2 p020 m3 p002 m1 p110 m1 p101 2 2 0 ( m1 m 2 ) p110 1 1 m1 p200 m2 p020 m3 p011 m3 p101 2 2 0 ( m1 m3 ) p101 1 1 m1 p200 m2 p002 m2 p011 m2 p110 2 2 p200 p020 p002 0 ,and - 45 - p101 p110 p011 1 But for not an unusually large network, say N 50 , K 10 , N K 1 59 10 there are system states, 1.26 10 N 50 1.26 1010 equations, if all states are possible. It is impossible even for modern large-scale computer (A formidable work!). This is why the product form solutions for Jackson network still are valuable. - 46 -