Optimal Routing Policy in Two Deterministic Queues

Transcription

1 INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE Optimal Routing Policy in Two Deterministic Queues Bruno Gaujal Emmanuel Hyon N 3997 Septembre 2000 THÈME 1 ISSN ISRN INRIA/RR FR+ENG apport de recherche

2

3 Optimal Routing Policy in Two Deterministic Queues Bruno Gaujal, Emmanuel Hyon Ý Thème 1 Réseaux et systèmes Projet TRIO Rapport de recherche n 3997 Septembre pages Abstract: We consider the problem of routing customers to one of two parallel queues where inter-arrival times and service times are deterministic. We provide an explicit formula for the average waiting time of the customers sent to one of the queues when the routing policy is an upper mechanical word. This formula is based on a special continued fraction decomposition of the service time in the queue. Using this formula we provide an algorithm computing the optimal routing policy for two queues. In general, this policy is an upper mechanical word with a rational ratio, and hence is periodic. Key-words: Mechanical Words, Continued Fractions, Deterministic Queues LORIA, 615 route du jardin botanique, BP 101, F Villers-les-Nancy, France. Bruno.Gaujal@loria.fr. Ý LORIA, 615 route du jardin botanique, BP 101, F Villers-les-Nancy, France. Emmanuel.Hyon@loria.fr. Unité de recherche INRIA Lorraine LORIA, Technopôle de Nancy-Brabois, Campus scientifique, 615, rue du Jardin Botanique, BP 101, Villers-Lès-Nancy (France) Téléphone : Télécopie :

4 Politique de routage optimale de deux files déterministes Résumé : Dans cet article nous étudions le routage de clients dans deux files d attente déterministes parallèles, les temps d interarrivées étant supposés déterministes. Nous donnons une formule explicite de la moyenne du temps d attente des clients introduits dans l une des deux files quand la politique de routage est un mot mécanique. Cette formule repose sur une décomposition en fraction continue particulière du temps de service. A partir de cette formule nous obtenons un algorithme qui calcule la politique de routage optimale dans les deux files. Généralement cette politique est un mot mécanique périodique et la proportion de clients envoyés dans chaque file est rationnelle. Mots-clés : Mots mécaniques, Fractions continues, Files déterministes

5 Optimal Routing Policy in Two Deterministic Queues 3 1 Introduction A classic problem in the literature of the control of queues is the determination of the optimal routing policy for parallel queues. When the queues are homogeneous (i.e. all customer service times are identically distributed and independent) and the service discipline is FIFO, then the optimal policy is often the join the shortest queue policy, as shown in [5], for many cost functions. When the service times in the queues have different distributions, the optimal policy for two parallel queues with exponential service times was studied in [4] using dynamic programming. Threshold policies were proved optimal. The non-markovian case with no state information was studied in [2] where it was shown that the optimal policy in terms of average waiting time must be a mechanical word, although no explicit computation of an optimal policy was actually provided in the heterogeneous case. In the deterministic case (both the arrivals and the service times are constant deterministic variables), the optimal policy was computed in [8] when the system is fully loaded. Here, we consider two deterministic FIFO queues with infinite buffers and arbitrary constant arrival and service times (hence the load is arbitrary). As proved in [2], the optimal policy is a mechanical word, and we are able to compute explicitly the slope of this optimal word, which is a rational number as long as the system is not fully loaded. Hence, the optimal policy is always periodic even when the parameters (service times and arrival times) are irrational numbers. In order to obtain this result, we start in Section 2 by introducing upper and lower mechanical words. In Section 3, we define a new continued fraction decomposition which helps us for identifying special factors of upper mechanical words. In Section 4, we give an explicit formula for the average waiting time in one deterministic queue when the arrival process is an upper mechanical word. This function is continuous but not differentiable at certain rational points called jumps in the following. It is increasing and concave between jumps. Finally, in Section 5, we consider the case of two queues. Using the properties of the average waiting time in one queue, we can show that the optimal routing policy is given by a jump in one of the two queues, and hence is periodic, as long as the system is not fully loaded. An algorithm is provided that computes the optimal jump in finite time (and hence the optimal policy). In the last section, several examples are studied in detail in order to illustrate the strange behavior of the optimal policy in such a simple system (or so it seems). 2 Mechanical words Let ¼ be the alphabet. The free monoid is the set of the finite words on. An infinite word is an element of Æ. Definition 1 (Slope). The slope of a finite nonempty word Û is the number : «Ûµ Û Û where Û is the number of letters equal to one in Û and Û is the length of Û. Let Û Ò Ò be the prefix of length Ò of an infinite word Û. If the sequence «Û Ò µ converges when Ò the limit is called the slope of Û. For a real number Ü, we denote by Ü (resp. Ü) the largest (resp. smallest) integer smaller (resp. larger) than Ü. Definition 2 (Mechanical word). The upper mechanical word with slope «is the infinite word Ñ «where the Ò Ø letter, with Ò ¼, is : Ñ «Òµ Ò µ «Ò «(1) The lower mechanical word with slope «is the infinite word Ñ «where the Ò Ø letter, with Ò ¼, is : Ñ «Òµ Ò µ «Ò «(2) RR n 3997

6 4 B. Gaujal & E. Hyon The characteristic word of slope «is the infinite word «where the Ò Ø letter, with Ò ¼, is : «Òµ Ò µ «Ò µ «Lemma 3. If «is a rational number («Ô Õ ), then Ñ «Ñ ««are periodic of period Õ. If «is an irrational number, then Ñ «Ñ ««are all aperiodic. Proof. This result is proved for example in [6]. In the following, by a slight abuse of notation when an infinite word Û is periodic we also denote by Û its shortest period. Example 4 (Graphical interpretation). The terminology comes from the following graphical interpretation. For example let «, the lower mechanical word is ¼¼¼¼, the upper mechanical word is ¼¼¼¼ and the characteristic word is ¼¼¼¼. Consider the straight line with equation Ý «Ü, and consider the points with integer coordinates just below the line : È Ò Ò Ò«µ, the ones just above the line : ÈÒ ¼ Ò Ò«µ and the points È ¼¼ Ò Ò Ò µ«µ. The points È Ò form a representation of the lower mechanical word in the following sense : two consecutive points are joined by either an horizontal straight line segment,if Ò «Ò«¼, or a diagonal, if Ò µ«ò«. Similarly the points ÈÒ ¼ are a representation of the upper mechanical word and the points ÈÒ ¼¼ are a representation of the characteristic word (see figure 1). representation of upper mechanical word straight line : y=3 x / 7 representation of lower mechanical word characteristic word Figure 1: Mechanical words associated with the line Ý Ü. Remark 5. Since Ò«Ò«except when Ò«is an integer, one has Ñ «Òµ Ñ «Òµ when Ò«and Ò µ«are not integer numbers. When «is an irrational number, since Ñ «¼µ and Ñ «¼µ ¼ then Ñ «¼ «and Ñ ««. When «is a rational number Ñ «is a shift of Ñ «(i.e. there exists a integer number such that Ñ «Ò µ Ñ «Òµ for all Ò). Lemma 6. Let «be a real number, then the maximum number of consecutive 0 in Ñ «is «the minimum number of consecutive 0 in Ñ «is «. and Proof. If «is an integer number it can be checked that Ñ «is the word composed by one followed by «consecutive 0. Suppose now that «is not an integer number. Let such that Ò µ«ò«¼ and Ò µ«ò«that means we have consecutive letters of Ñ «equal to ¼. Let Ò «Ò «Ò. Let Ò such that Ò µ««. Note that such a Ò exists since for a given we can find Ò such that Ò. We have Ò µ«ò«¼ µ µ«ò«ò«µ µ«ò Ò µ«ò«µ µ«ò«ò«µ µ«ò Hence Ò µ«µ Ò µ«. Then «µ «INRIA

7 Optimal Routing Policy in Two Deterministic Queues 5 3 Expansions in continued fractions 3.1 Classical expansion in continued fraction In this part we will exhibit the link between expansion in continued fraction and the prefixes of characteristic words as presented in [6] and [7]. Let ¼ be a real number. The computation of its expansion in a continued fraction is given by : µ ¼ Ñ Ñ Ò Ñ Ò Ò Ñ Ò Ò µ (3) Ò If is a rational number, the process stops after a finite number of steps when ( Ò A simple continued fraction ¼ Ñ Ñ Ñ Ò Ñ ). Ò ¼ is symbolically written in the form ¼ Ñ Ñ Ñ Ò. Let Ñ Ñ Ñ Ò µ be a sequence of integers, with Ñ ¼ and Ñ Ò ¼ for Ò. To such a sequence we associate a sequence Ò µ Ò µ of finite words defined by ¼ ¼ Ò ÑÒ Ò Ò Ò µ This sequence Ò µ Ò is called the standard sequence associated with the sequence Ñ Ò µ Ò. Proposition 7. Let ¼ Ñ Ñ Ñ Ò be the continued fraction expansion of some irrational with ¼, and let Ò µ Ò be the standard sequence associated to (Ñ Ñ ). Then every Ò is a prefix of the characteristic word and Proof. This result is proved in [6]. ÐÑ Ò Ò 3.2 Mechanical expansion in continued fractions The tight relation between the continued fraction of and the characteristic word does not extent easily to the mechanical words. In the following, we will define a special expansion in continued fraction (called mechanical continued fraction expansion). This new continued fraction allows us to find the decomposition in special factors of upper mechanical words. The construction of the mechanical continued fraction expansion of a number «with ¼ «is given by : «Ð «Ð ««Ò Ð Ò «Ò Ð Ò «Ò µ (4) Ò When «is a rational number the construction finishes after a finite number of steps when «Ò Ð Ò. When «is an irrational number the construction is infinite and we obtain the infinite mechanical continued fraction expansion of «. An infinite mechanical continued fraction expansion is written in the form Ð Ð Ð Ð Ò Ð Ò. A finite mechanical continued fraction is written in the form Ð Ð Ð Ò Ð Ò. A partial mechanical continued fraction of a number «is written in the form Ð Ð Ð Ò «Ò. RR n 3997

8 6 B. Gaujal & E. Hyon Let us denote by Å the set of all upper mechanical words. Denote by Å the set of upper mechanical words with slope «such that µ «. Let us now introduce the morphism ³ : Å Æ. ³ Þ Ð ß ¼ ¼ Þ Ð ß ¼ ¼ ¼ If «then the word ³ Ñ «µ is reduced to the letter one. Introduce the morphism We now show that has its values in Å. Å Æ Ñ «³ «Ñ «µ Lemma 8. Let Ñ «be the upper mechanical word of slope «, then Ñ «µ is the upper mechanical word of slope ««µ. Proof. If «is an integer number Ñ «µ is reduced to the word, then we have the result since ««µ. Recall that Ñ «Òµ is defined by Formula (1). An integer is the index of the Ø occurrence of the letter one in Ñ «if Ñ «¼µ Ñ «µ contains letters one and Ñ «¼µ Ñ «µ contains letters one. This means «µ «These equalities imply «. Define now the function, : Å Æ. Þ Ð ß ¼ ¼ Þ Ð ß ¼ ¼ ¼ Let Ð «and let Û Ð Ñ «µ. The sequence Û µµ ¼ of letters of Û can be computed by Û µ Ð. Then Û µ µ«ð µ «Ð, hence Û is the lower mechanical word of slope «Ð Define now the function such that ¼ ¼ The function transforms a lower mechanical word of slope «into an upper mechanical word of slope «and conversely. Let Û ¼ Ûµ, Û ¼ is the upper mechanical word of slope «Ðµ. It can be checked that Æ, hence Ñ «µ is the upper mechanical word of slope ««µ. The following corollary shows the relation between Ñ «µ and the mechanical continued fraction of Ñ «. Corollary 9. Let «¼ «be any given real number such that «Ð Ð Ð Ò Ð Ò «Ò, with Ò. Let Û Ñ «µ. Then «Ûµ Ð Ð Ð Ò Ð Ò «Ò. Proof. By Equations (4), we have «Ð «µ 8, we obtain and «Ð Ð Ð Ò Ð Ò «Ò. By Lemma Hence «Ûµ Ð Ð Ð Ò Ð Ò «Ò. «Ûµ ««µ «Ð µ «INRIA

9 Optimal Routing Policy in Two Deterministic Queues 7 This will allow us to use induction on the number of terms in the mechanical continued fraction in the following. Theorem 10 ((x,y)-factor decomposition). Let «¼ «be any given real number with «Ð Ð Ð Ò. Define two sequences, Ü «µ ¼ and Ý «µ ¼, computed by : Ü ¼ «µ Ý ¼ «µ ¼ Ü «µ Ü «µ Ý «µµ Ð Ý «µ Ü «µ Ý «µµ Ð with Then the upper mechanical word Ñ «can be factorized only using the two factors Ü «µ and Ý «µ. These two sequences are called (x-y)-factor decomposition sequences associated with the mechanical expansion of «. Proof. We will study finite sequences Ü «µ ¼Ò, Ý «µ ¼Ò associated with the partial mechanical continued fraction expansion of «and prove the result by induction. Step 1. We have «Ð «. By Lemma 6, it is checked that Ñ «can be factorized only using Ü and Ý. Step n. We have «Ð Ð Ð Ò ÐÔ Ò. Let Û Ñ «µ, by Corollary 9, «Ûµ Ð Ð Ð Ò Ð Ò «Ò. Let Ü «Ûµµ ¼Ò and Ý «Ûµµ ¼Ò be the two (x-y)-factor decomposition sequences associated with the partial mechanical expansion of «Ûµ. By induction hypothesis on the number of terms in the partial expansion Û can be factorized only using Ü «Ûµµ and Ý «Ûµµ. Introduce now the sequences Ü ¼ and ݼ such that Ü ¼ ¼ ݼ ¼ ¼ ³ Ð Ü ¼ µ Ü «Ûµµ ³ Ð Ý ¼ µ Ý «Ûµµ Then Û Ñ «µ can be factorized only using ³ Ð Ü ¼ µ and ³ Ð Ý ¼ µ, ¼. Therefore Ñ «can be factorized only using Ü ¼ and ݼ, since Ñ «µ ³ Ð Ñ «µ. Note now that the sequences Ü ¼, ݼ are the (x-y)-factor decomposition sequences associated with Ð Ð Ð Ò. Lemma 11. Let «¼ «be any given rational number with «Ð Ð Ð Ò Ð Ò. Then Ü Ò «µ Ñ «. Proof. We also use here an induction on the number of terms in the partial mechanical expansion. Step 1. Considering that «Ð then «is a rational number with «Ð. Hence Ñ «¼ ¼, where the number of consecutive 0 is «Ð. Step n. We have «Ð Ð Ð Ò Ð Ò, therefore Û Ñ «µ has a slope «Ûµ Ð Ð Ò. Using the induction hypothesis we obtain Û Ü Ò «Ûµµ. Since Ü Ò «Ûµµ ³ Ð Ü Ò «µµ, then Ñ «Ü Ò «µ. Remark 12. Since Òµ Ñ «µ ³ ÐÒ Æ Æ ³ Ð Æ ³ Ð Ñ «µ, then Ü Ò «µ and Ý Ò «µ are the only factors of Ñ «, such that Òµ Ü Ò «µµ and Òµ Ý Ò «µµ ¼. Remark 13. Since Ò ¼, Òµ Ñ «µ is an upper mechanical word and since all upper mechanical words begin with one, then the sequence Ü «µ is a sequence of prefixes of Ñ «verifying ÐÑ Ü «µ Ñ «Example 14. Here «. The mechanical expansion of «using Equations (4) is :. Ñ ¼¼¼¼¼¼¼¼¼¼¼ Ñ µ ¼¼¼, µ Ñ µ ¼¼¼ and µ Ñ µ ¼. Using the (x-y)-factor decomposition we obtain Ü ¼ and Ý ¼¼ ; Ü Ü and Ý Ü Ý ; Ü Ü Ý and Ý Ü Ý µ with finally Ü Ü Ý. We can check Theorem 10 and Lemma 11. Ü Þ Ð ß Þ Ð ß Ü ÞÐß ÞÐß ÞÐß ÞÐß ÞÐß ÞÐß ÞÐß Ñ ¼ ¼ ¼¼ ¼ ¼ ¼¼ ¼ ßÞÐ ßÞ Ð ßÞÐ ßÞ Ð Ü Ü Ý Ý Ü Ü Ü Ý Ý Ý Ü Ý ÞÐß ¼¼ ßÞ Ð Ý ßÞ Ð Ü RR n 3997

10 8 B. Gaujal & E. Hyon Theorem 10 and Lemma 11 underline the connection between the construction of factors of an upper mechanical word and the mechanical continued fraction expansion of its slope. This connection could be seen like an equivalent for upper mechanical words of the relation between continued fraction expansion and characteristic words presented in Proposition (7). In the following, the Ø factors of the word Ñ «will be denoted Ü and Ý when no confusion is possible. 4 Average waiting time in a single queue The aim of this section is to compute and to study the average waiting time of customers, Ï Ë Ûµ in a./d/1//fifo queue with a constant service time, Ë, and with constant inter-arrival times. Using an appropriate time scale, the inter-arrival times can be chosen to be equal to one. The input sequence Û is such that Û Òµ is ¼ if no customer enter the queue at slot Ò and Û Òµ is if one customer enters the queue at slot Ò. In the following, we will only consider an input sequence which is the upper mechanical word of slope «, Ñ «. Let «¼ «be the ratio of customers sent in the./d/1//fifo queue. The stability condition of the queue is «Ë (5) 4.1 Jumps We now show that the computation of the average waiting time Ï Ë Ñ «µ tightly depends on the factorization of Ñ «in (x-y)-factors of Ë. For that we will exhibit special rational numbers. Lemma 15. Let Ô Ð Ð Ð Ò Ð Ò Ð Ò be a rational number. -i) For all real number «which partial mechanical expansion is Ð Ð Ð Ò Ð Ò «Ò, such that Ð Ò is the smallest integer satisfying «Ò Ð Ò Ð Ò Then Ü Ò «µ Ü Ò Ôµ Ý Ò Ôµµ ÐÒ and Ý Ò «µ Ü Ò Ôµ. -ii) For all real number which partial mechanical expansion is Ð Ð Ð Ò Ð Ò Ò with Ð Ò Ð Ò Ò then Ü Ò µ Ü Ò Ôµ Ý Ò Ôµµ and Ý Ò µ Ü Ò Ôµ Ý Ò Ôµµ with Ð Ò -iii) For any number which partial mechanical expansion is Ð Ð Ð Ò Ð Ò Ò, then Ò is the proportion of Ü Ò µ Ü Ò «µ Ü Ò Ôµ in Ñ (i.e. the number of Ü Ò µ divided by the number of Ü Ò µ added with the number of Ý Ò µ in Ñ «). Proof. Note that Ð Ò Ð Ò µð Ò. Hence the Equations (4) lead to Ô «. -i) For all ¼ Ò, we have Ü Ôµ Ü «µ and Ý Ôµ Ý «µ, since the mechanical expansions are equal until the Ò Ø coefficient. Compute Ð Ò the smallest integer such that that implies «Ò Ð Ò Ð Ò Ð Ò «Ò µ Ð Ò «Ò Hence the mechanical expansion of «is Ð Ð Ð Ò Ð Ò µ «Ò. Therefore Ü Ò «µ Ü Ò Ôµ Ý Ò Ôµµ ÐÒ and Ý Ò «µ Ü Ò Ôµ Ý Ò Ôµµ ÐÒ Ü Ò Ôµ. INRIA

11 Optimal Routing Policy in Two Deterministic Queues 9 -ii) ÐÒ Ð Ò Ò implies Ð Ò Ò therefore the number of consecutive Ý Ò µ in Ü Ò µ is larger than Ð Ò. Since Ü Ò Ôµ Ü Ò µ and Ý Ò Ôµ Ý Ò µ then we proved the result. -iii) By Remark 12, Ü Ò µ and Ý Ò µ are the only factors of Ñ, such that Òµ Ü Ò µµ and Òµ Ý Ò µµ ¼. Hence the composition in Ü Ò µ, Ý Ò µ of Ñ is the composition in and ¼ of Òµ Ñ µ. By Lemma 8 and Corollary 9, Òµ Ñ µ is an upper mechanical word of slope Ò. Therefore using Definition 1, Ò is the proportion of Ü Ò in Ñ. The previous Lemma says that in all mechanical words with slope «Ô Ð Ð Ð Ò Ð Ò Ð Ò, the sequence Ü Ò ÔµÝ Ò Ôµ ÐÒ never appears. But as soon as the slope is greater than Ô then the sequence Ü Ò ÔµÝ Ò Ôµ ÐÒ appears. Intuitively, when you increase the slope the number of ones increases and the number of zero decreases, hence the minimum number of consecutive Ý Ò decreases until Ü Ò ÔµÝ Ò Ôµ ÐÒ appears. Definition 16 (jumps). Denote the mechanical continued fraction expansion of Ë by Ð Ð Ò. Define the rational number Ö Ë µ by Ö Ë µ Ð Ð Ð Ð. The number Ö Ë µ is called the Ø jump of Ï Ë. When Ë is a rational number, then Ë Ð Ð Ð Æ and we define the last jump Ö Æ Ë µ by Ö Æ Ë µ Ð Ð Ð Æ. Note that the jumps form a sequence of rational numbers increasing towards Ë. When Ë is rational, the number of jumps is finite and the last jump equals Ë. Let us define Ò and Õ to be the two relatively prime integers such that Ö Ë µ Ò Õ. Therefore by Definition 1, Ò Ñ Ö Ë µ and Õ Ñ Ö Ë µ. For example, if Ö Ð Ð, then it can be checked that Ò Ð and Õ Ð Ð µ Ð. From now on Ö Ë µ will be denoted by Ö when no confusion is possible. Lemma 17. The number Ö is the rational number with the smallest denominator in Ö Ë. Proof : Let «Ð Ð Ð «be any given rational number in Ö Ë. By Equations (4) we have «µ Ð. Hence by Lemma 15, Ü Ý µ Ð is a factor of Ñ «. That means Ñ «Ü Ý µ Ð. Since Ü Ý µ Ð is the mechanical word Ñ Ö then Ü Ý µ Ð Õ. Considering that Ñ «is the denominator of «ends the proof. Remark 18 (Connection with classical continued fractions). The rational number, obtained by keeping only terms in a classical expansion in continued fractions is called the Ø convergent. The even convergents Ò of a simple continued fraction form an increasing sequence. Hence if Ñ ¼ Ñ Ñ denote the classical continued fraction of Ë, then the even convergents satisfy Ò Ñ ¼ Ñ Ñ Ò Ë. If Ò Ñ ¼ Ñ Ñ Ò is an even convergent of Ë, then the rational numbers Ñ ¼ Ñ Ñ Ò Ò with Ò Ñ Ò are called intermidiate convergents. Note now that the sequence of intermediate convergents and even convergents is the sequence of jumps Ö Ë µ, since this sequence increases towards Ë and each term of this sequence is the rational with the smallest denominator in the interval formed by the preceeding intermediate convergent and Ë and since Ö. However it seems hard to use the intermediate convergents to compute the average waiting time as it is shown further. 4.2 Formula of the average waiting time We will see that the factors Ü tend to increase the load in the queue whereas the factors Ý tend to decrease the load. The decomposition of Ñ Ö in (x-y) factors of Ë allows us to compute recursively the average waiting time of the input sequence Ñ Ö : Ï Ë Ñ Ö µ. From that we can compute Ï Ë Ñ «µ for any rational number «Ö «Ö and finally we extend it to any irrational number «. RR n 3997

12 10 B. Gaujal & E. Hyon In the following, the workload denotes the amount of service (in time units) remaining to be done by the server. Let denote the increase of the workload after an input sequence equal to Ü. Let denote the maximal possible decrease of the workload after an input sequence equal to Ý. Lemma 19. If the initial workload is equal to zero then using Ü as input sequence the workload during Ü is never null and remains non negative at the end of the sequence. If the initial workload is equal to zero then using Ý as input sequence the workload is never null during Ý until its last letter (which is always a ¼) and is null at the end of the sequence. Proof. The proof holds by induction on. Step 1. Assume Ë is not an integer, we have Ð Ë Ð. Therefore Ë Ð ¼ Ë Ð and the result is proved for step one since the workload after Ü is equal to Ë Ð and the workload after Ý is ÑÜ Ë Ð ¼µ ¼. Moreover Equations (4) yield «and «. When Ë is an integer, we have Ë Ð ¼. Step 2. By Theorem 10, Ü is composed by Ü followed by Ð µ consecutive Ý and Ý is composed by Ü followed by Ð consecutive Ý. Hence to prove the result we have to show that ÑÜ Ð µ : «Ð µ «µ ¼ «Ð µ «µ ¼ «Ð µ «µ ¼ this last inequality implying the nullity of the workload. By Lemma 15, Ð is the smallest integer satisfying «Ð Ð µ, then for all such that ÑÜ Ð µ we have Ð Ð «Ð Ð Since Ð µ Ð µ Ð µ Ð µ, the three inequalities are satisfied. It can be checked that «Ð µ «µ «µ«and «Ð µ «µ «µ «µ. Therefore «µ«and «µ «µ. Step i+1. Suppose that «µ «µ «µ and «µ «µ«. We have to show that ÑÜ Ð µ : «µ «µ «Ð µ «µ ¼ (6) «µ «µ «Ð «µ ¼ (7) These inequalities are satisfied indeed, since Ð is the smallest integer such that «Ð Ð µ. Moreover, by the induction hypothesis, during the sequence Ý the workload is positive until the last ¼, when the initial workload is null. Consider now the last factor Ý in Ý, its initial workload is positive by Inequality (6). If the workload is never null during an input sequence with an initial workload equal to zero, then the same result holds with a positive initial workload. Hence the workload is never null during the last Ý and also during Ý, up to its last letter. Using «Ð µ «µ «µ«and «Ð µ «µ «µ «µ we obtain That finishes the proof. «µ «µ «µ and «µ «µ«(8) Remark 20. When Ë Ð Ð Ð Æ is a rational number, «Æ Ð Æ which is equivalent to «Æ Ð Æ µð Æ. Then «Æ Ð Æ µ «Æ µ ¼. This yields Æ ¼ and Æ ¼. Hence the workload during Ü Æ is positive until the last epoch of the admission sequence Ü Æ Ëµ Ü Æ Ö Æ µ, and null at the end. INRIA

13 Optimal Routing Policy in Two Deterministic Queues 11 We give now a direct consequence of Lemma 19. The workload workload can be recursively computed by : and the maximal decreases of the Ð µ Ð µ with Ë Ð Ë Ð (9) Example 21 (Average waiting time with S = 51/20 and Ñ ). Let Ë ¼. The partial mechanical expansion of order three of ¼ is. The (x,y)-factor decomposition gives Ü ¼, Ý ¼¼, Ü ¼¼¼, Ý ¼¼¼¼¼, Ü Ü and Ý ¼¼¼¼¼¼¼¼. Note that Ö ¼µ and that Ñ Ý. The Equation (9) leads to ¼ ¼ ¼ ¼ ¼ ¼ ¼¼ ¼ ¼ ¼ ¼ ¼ The figure 2 represents the workload during a sequence Ý. The quantity is represented by a, by b, by c, by d and by e a b c d e Figure 2: Representation of the workload We have Ñ Ö Ü Ð Ý, (Ñ ÖÆ Ü Æ Ð Æ µý Æ when Ë Ð Ð Ð Ð Æ ) and Ü Ý Ý. The recursive computation of Ò Ñ Ö is given by : Ò ¼ ¼ (10) Ò (11) Ò Ð µò Ò (12) Ò Æ Ð Æ Ò Æ Ò Æ if Ë Ð Ð Æ (13) This allows us to compute the average waiting time of Ï Ë Ñ Ö µ recursively. The sum of the waiting times of customers admitted during the sequence Û when the first customer of the sequence waits for a time Ø is denoted by Ã Ø Ûµ. We also denote the sum of the waiting times of Ñ Ö over one period by Ã Ø and ü by Ã. Note first, that we can focus on one period since the workload after Ñ Ö is null. Lemma 22 (Formula of Ï Ë Ñ Ö µ). The average waiting time of Ñ Ö is : Ï Ë Ñ Ö µ Ã Ò RR n 3997

14 12 B. Gaujal & E. Hyon with à ¼ à РРµ Ð µ (14) or if Ë Ð Ð and à Рµ Ð µ Ð µ (15) à Рµã Ã Ò Ð Ò µ Ò Ð Ð µ (16) and if Ë Ð Ð Æ Ã Æ Ð Æ Ã Æ Ã Æ Ò Æ Ð Æ µ Ò Æ µ Æ Ò Æ Ð Ð Æ Æ µ Æ (17) Proof. The minimum number of consecutive zeros in Ñ Ö is equal to Ð. This means that the inter-arrival time in the queue is equal to Ð and since Ë Ð none of the customer has to wait, therefore à ¼. Recall that we have by Definition 16 Ñ Ö Ü Ý µ Ð. The waiting time of a customer is the workload at the epoch of its arrival. By Lemma 19 the first customer admitted does not wait, on the other hand the next Ð customers will have to wait. Moreover the workload during Ñ Ö is positive until the last 0 hence the waiting time of the second customer is, the one of the third customer is, the one of the Ø is µ and the one of the last customer, which is the Ð µ Ø, is Ð µ. Summing we obtain à РРµ Ð µ. Let Û be an admission sequence during which the workload is never null if the initial workload is equal to zero. When the first customer of Û waits for a time Ø, the waiting time of a customer admitted during Û is its waiting time when the first customer does not wait, increased by Ø. Therefore Ã Ø Ûµ Û Øµ à ¼ Ûµ. We have Ñ Ö Ü Ý µ Ð. Let us compute the waiting times of the first customers of the Ð consecutive Ý. The waiting time of the first customer of the first Ý is. The waiting time of the first customer of the second Ý is. The waiting time of the first customer of the Ø Ý is µ. The waiting time of the first customer of the last Ý (the Ð Ø ) is Ð µ. We decompose Ã. à à ¼ Ü µ à à à µ à ¼ Ü µ Ò Ã Ò µ à à ¼ Ü µ Ð Ã Ò Ð Ð µ Ð µ since Ý Ü Ý, à à à à Рµ Ò Ð µ µ Ã Ð Ã Ò Ð Ð µ Ð µ Ã Ò Ã µ Ð Ã Ò Ð Ð µ Ð µ When Ë Ð Ð Ð Æ, the Æ-th jump is a secial case since Ñ ÖÆ Ë µ Ñ Ë is Ü Æ and the number of consecutive Ý Æ following Ü Æ is Ð Æ. Theorem 23 (Formula of Ï Ë Ñ «µ for rational ratio). Let Ö be the Ø jump of Ï Ë Ñ «µ, and Ö the Ø jump of Ï Ë Ñ «µ. Let «Ð Ð «be a rational number such that «Ö Ö µ. Then the average waiting time of Ñ «is Ï Ë Ñ «µ «µã «Ð «µ à «µò «Ð «µ Ò (18) If «is a rational number such that «Ö, then Ï Ë Ñ «µ ¼. INRIA

15 Optimal Routing Policy in Two Deterministic Queues 13 Proof. By Theorem 10, Ñ «can be decomposed only using Ü «µ and Ý «µ. Since by Equations (4) Ð Ð «then we obtain the composition of Ü «µ and Ý «µ with Lemma 15: Ü «µ Ü Ö µ Ý Ö µµ and Ý «µ Ü Ö µ Ý Ö µµ with Ð µ Let us rewrite these two equalities into : Ü «µ Ü Ö µ Ý Ö µµ Ð Ý Ö µµ ¼ and Ý «µ Ü Ö µ Ý Ö µµ Ð Ý Ö µµ ¼ with ¼ ¼. We have Ü «µ Ñ Ö Ñ Ö µ ¼ and Ý «µ Ñ Ö Ñ Ö µ ¼. Since by Lemma 19 and Remark 20 the workload after Ñ Ö is always null, then the workload after an admission sequence Ü «µ or Ý «µ is also null. Therefore the workload after one period of Ñ «is null and we simply have to focus on one period of Ñ «. Computing à ¼ Ü «µµ and à ¼ Ý «µµ gives à ¼ Ü «µµ à ¼ à and à ¼ Ý «µµ à ¼ µã We write «ÒÕ, «being rational. The number of Ñ Ö in Ñ «is the number of Ü Ë µ Ü Ö µ in Ñ «. By Lemma 15 the total number of Ü Ë µ in Ñ «is equal to the total number of ones in µ Ñ «µ. Since ÒÕ is the slope of µ Ñ «µ and since Ñ «and µ Ñ «µ are finite words. Then by Definition 1 the total number of Ü Ë Ò µ in Ñ «is Õ µõ. Using a similar argument the total number of Ñ Ö in Ñ «is Ò Õ µõ. The number of Ñ Ö which do not belong to Ñ Ö is equal to Õ Ò Õ µ Ð Ò Õ µõ. Therefore Ã Ñ «µ Õ Ò Õ µã Õ Ò Õ µ Ð Ò Õ µ à Compute now Ò «the total number of ones in Ñ «. This number is the number of Ñ Ö multiplied by the number of ones in Ñ Ö added with the number of Ñ Ö which do not belong to Ñ Ö multiplied by the number of ones in Ñ Ö : We obtain Ò «Õ Ï «Ëµ Ò Õ µò Õ Ò Õ µ Ð Ò Õ µã Ò Õ Ð Ò Õ µò Ò Õ Ð Ò Õ µ Ò Ò Õ µµ Ã Ò Õ µµ Ò When «Ö, by Lemma 6 the minimum number of consecutive 0 in Ñ «is greater than Ð. Therefore the inter-arrival times of Ñ «are greater than Ð, hence none of the customers has to wait and Ï Ë Ñ «µ ¼. We will now extend this result to irrational numbers. First, we will define the average waiting time of an infinite word Û. Let Û Ò be the prefix of length Ò of Û, then if the limit exists. Ï Ë Ûµ à ¼ Û Ò µ ÐÑ Ò Û Ò Theorem 24 (Formula of Ï Ë Ñ «µ for irrational ratio). Let «Ð Ð «be an irrational number such that «Ö Ö µ, where Ö is the Ø jump of Ï Ë Ñ «µ, and Ö is the Ø jump of Ï Ë Ñ «µ. The average waiting time of Ñ «is also given by the Formula (18) that is : Ï Ë Ñ «µ «µã «Ð «µ à «µò «Ð «µ Ò RR n 3997

16 14 B. Gaujal & E. Hyon Proof. Let Ò be an integer. Let Ñ «µ Ò be the prefix of length Ò of Ñ «. Let Ñ «µ Ò be the greatest prefix of Ñ «uniquely composed by factors Ü «µ and Ý «µ which length is smaller than Ò, and let Ñ «µ Ò be the smallest prefix uniquely composed by factors Ü «µ and Ý «µ which length is greater than Ò. We obtain à ¼ Ñ «µ Ò µ ü Ñ «µ Ò µ ü Ñ «µ Ò µ Ñ «µ Ò Ñ «µ Ò Ñ «µ Ò The prefix Ñ «µ Ò is composed by Ñ «µ Ò followed by either Ü «µ or Ý «µ. As shown above, the workload after any sequence Ü «µ and Ý «µ is null. Hence the workload after Ñ «µ Ò and after Ñ «µ Ò is also null. Since the factor Ý «µ is longer than the factor Ü «µ, then Ü «µ Ý «µ and à ¼ Ü «µµ à ¼ Ý «µµ. That yields à ¼ Ñ «µ Ò µ à ¼ Ñ «µ Ò µ à ¼ Ý «µµ and Ñ «µ Ò Ñ «µ Ò µ Ý «µ. Therefore à ¼ Ñ «µ Ò µ Ñ «µ Ò Ã¼ Ñ «µ Ò µ ü Ñ «µ Ò µ ü Ý «µµ (19) Ñ «µ Ò Ñ «µ Ò Ý «µ Ñ «µ Ò Ñ «µ Ò Ñ «µ Ò Let us focus on à ¼ Ñ «µ Ò µ. For all Ò Ò, this sum is strictly positive. Moreover this sum, as shown in the previous proof, only depends on the number of Ý Ö µ and remaining Ý Ö µ. The number of Ý Ö µ and remaining Ý Ö µ is given by the slope of µ Ñ «µ Ò µµ. Hence and à ¼ Ñ «µ Ò µ µ Ñ «µ Ò µµ «µ Ñ «µ Ò µµ à µ Ñ «µ Ò µµ «µ Ñ «µ Ò µµ Ñ «µ Ò µ Ñ «µ Ò µµ «µ Ñ «µ Ò µµ Ò µ Ñ «µ Ò µµ «µ Ñ «µ Ò µµ Ð «µ Ñ «µ Ò µµ à Р«µ Ñ «µ Ò µµ Ò Finally we have à ¼ Ñ «µ Ò µ Ñ «µ Ò «µ Ñ «µ Ò µµã «µ Ñ «µ Ò µµ Ð «µ Ñ «µ Ò µµ à «µ Ñ «µ Ò µµò «µ Ñ «µ Ò µµ Ð «µ Ñ «µ Ò µµ Ò Note now that the number Ñ «µ Ò depending on Ò, when Ò, Ñ «µ Ò and also Ñ «µ Ò. Definition 1 and Lemma 15 give This implies ÐÑ «µ Ñ «µ Ò µµ ÐÑ «µ Ñ «µ Ò µµ «Ò Ò Ã ¼ Ñ «µ Ò µ ÐÑ «µã «Ð «µ à (20) Ò Ñ «µ Ò «µò «Ð «µ Ò Since Ý Ö µ is a finite factor then Ý Ö µ and Ã Ý Ö µµ are bounded. It comes à ¼ Ñ «µ Ò µ Ñ «µ Ò Ã ¼ Ñ «µ Ò µ ÐÑ ÐÑ Ò Ñ «µ Ò Ñ «µ Ò Ý «µ Ò Ñ «µ Ò Ã ¼ Ñ «µ Ò µ ü Ý «µµ à ¼ Ñ «µ Ò µ ÐÑ ÐÑ Ò Ñ «µ Ò Ñ «µ Ò Ò Ñ «µ Ò INRIA

17 Optimal Routing Policy in Two Deterministic Queues 15 Using Equation (19) and Equation (20) we obtain «µã «Ð «µ à à ¼ Ñ «µ Ò µ ÐÑ «µò «Ð «µ Ò Ò Ñ «µ Ò «µã «Ð «µ à «µò «Ð «µ Ò 4.3 Properties Here, we will give some properties of the function Ï Ë Ñ «µ. Most of them are required to compute the optimal ratio in the case of two queues. Lemma 25 (Continuity). The function «Ï Ë Ñ «µ is continuous. Proof. Let «Ð Ð Ò «Ò and Ð Ð Ò Ò be two points in the interval. Note that «and are both in the interval Ö Ò Ö Ò. If «goes to zero, then «Ò Ò also goes to zero. By Equation (18), Ï Ë Ñ «µ Ï Ë Ñ µ goes to zero. Using similar arguments, one can show that the function «Ï Ë Ñ «µ is infinitely differentiable on each interval Ö Ò Ö Ò, but it is not differentiable at the jumps. Lemma 26. The function Ã Ò Proof. We have is strictly increasing. Ã Ò Ã Ò Ò Ã Ò Ã Ò Ò Since, Ò Ò ¼ we have to focus on the sign of Ò Ã Ò Ã. Using Equations (12) and (16) yields Ò Ã Ò Ã Ò Ð µã Ò Ð Ò µ Ò Ð Ð µ Ã Ò Ð µã Ã Ò Ð Ò µ Ò Ð Ð µ Ò Ã Ò µã Ò Ð µã Ò Ã Ò Ã µ Ò Ò Ð Ò µ Ò Ð Ð µ reordering we obtain Ò Ã Ò Ã Ò Ã Ò Ã µ Ò Ð µ Ð µ Ò Ò Ò µ (21) On the first hand, by Lemma 19, we know that is strictly positive and is non negative and that and are non positive. On the other hand Equation (9) gives us Ð µ and Ð. This implies Ð, That means Ð, Ð ¼ Ð µ Ð µ ¼ RR n 3997

18 16 B. Gaujal & E. Hyon Since Ã Ò Ã Ã (à ¼) and since à Р¼, then à ¼, and by induction we have ¼ : Therefore ¼, Ð µ and since Ð µ Ò Ã Ò Ã ¼ (22) Ã Ò Ã Ò ¼ Remark 27. As proved in [8] if Ë is an irrational number then the average waiting time Ï Ë Ñ Ë µ. Hence the sequence Ã Ò goes to when goes to infinity. Lemma 28 (Monotonicity). If «Ö Ö, then the function Ï Ë Ñ «µ is strictly increasing in «. Proof. First we will show by induction on that the derivative function ««is non positive. Step 1. Recall that by Equations (4) we have : «Ð «µ this means ««Ð. Hence ««¼ Step. By Equations (4), «Ð «µ, then ««µ Ð. Therefore ««¼ (23) and using the induction hypothesis «««««¼ (24) «Second, using Equations (18) and (22), we have Since Ï Ë Ñ «µ Ò Ã Ò Ã µ ««µ Ò «Ð «µµ Ò ¼ (25) then Ï Ë Ñ «µ «Ï Ë Ñ «µ «««Ï Ë Ñ «µ «¼ Lemma 29 (Concavity). If «Ö Ö, then the function «Ï Ë Ñ «µ is strictly concave. Proof. First, we prove by induction on that the second derivative function ««is strictly positive. Step 1. It can be checked that Step. We obtain : «««««Ð «¼ ««««««INRIA

19 Optimal Routing Policy in Two Deterministic Queues 17 The Formula (23) and the induction hypothesis can be written respectively as : hence ««¼ and ««¼ Using formulas (18) and (25) it comes ««¼ (26) Ï Ë Ñ «µ «Ò Ã Ò Ã µ Ò Ò Ð µµ «µ Ò «Ð «µµ Ò Equations (4) and (12) give Ð and Ò ¼. This implies, with Equation (22), Ò Ã Ò Ã µ Ò Ò Ð µµ ¼. Let us also note that «Ð «µ= «µ«which yields «µ Ò «Ð «µµ Ò ¼. Therefore Ï Ë Ñ «µ «¼ (27) We have By inequalities 25, 26 and 27 we know Ï Ë Ñ «µ «««Ï Ë Ñ «µ Ï Ë Ñ «µ ««««««¼ Ï Ë Ñ «µ «¼ and Ï Ë Ñ «µ «¼ Then Ï Ë Ñ «µ «¼ Lemma 30 (Concavity, II). If «Ö Ö, then the function ««Ï Ë «µ is concave. Proof. First we show by induction on that the function «««««is nonnegative. Step 1. We check that Step. We obtain «««««««µ «Ð «µ «¼ «««««««««««««««««««««««««««««Using Equations (4), the Inequality (23) and the induction hypothesis we get : ««¼ «««µ «¼ and ««««¼ RR n 3997

20 18 B. Gaujal & E. Hyon therefore Therefore, we have «Ï Ë Ñ «µµ ««««««¼ (28) Ï Ë Ñ «µ «Ï Ë Ñ «µ ««Ï «Ë Ñ «µ «Ï Ë Ñ «µ «««««Ï Ë Ñ «µ ««««««««Ï Ë Ñ «µ «««Ï Ë Ñ «µ ««Inequalities (25), (27) and (28) give respectively Ï Ë Ñ «µ «¼ Ï Ë Ñ «µ «¼ and «««««¼ then «Ï Ë Ñ «µµ «¼ To illustrate the results obtained in this section let us take an example. Example 31 (Curve of Ï «µ). The mechanical expansion of is and Ñ is equal to ¼. The factor decomposition is : Ü and Ý ¼, then Ü Ü and Ý Ü Ý ¼, then Ü Ü and Ý Ü Ý ¼, then Ü Ü Ý ¼ Ñ. We have Ö µ, Ö µ, Ö µ and Ö µ, therefore Ò, Ò, Ò and Ò. Let us now compute Ã, we obtain à ¼, Ã, à and Ã. Hence we have : Ã Ò ¼ Ã Ò Ã Ò Ã Ò Recall that «Ð «Ð Ð «Ð Ð Ð «. Formula (18) leads to Ï Ë Ñ «µ ¼ for «Ï Ë Ñ «µ «Ï Ë Ñ «µ ««µ Ï Ë Ñ «µ «µ for ««for ««for ««The function Ï Ë Ñ «µ for all ¼ «, is shown in figure 3. 5 Average waiting time for two queues 5.1 Presentation of the Model From now on as presented in figure 4, we will study a system made of two./d/1//fifo queues, with constant service times Ë in queue and Ë in queue. It is assumed for convenience in the following that INRIA

21 Optimal Routing Policy in Two Deterministic Queues Figure 3: Curve of Ï Ñ «µ Ë is the smallest service time. The time unit is chosen such that inter-arrival time slots are constant and equal to one. When they arrive, the customers are routed to a queue where they wait to be treated in a FIFO order. The ratio of customers sent in queue is «and therefore, the ratio of customers sent in queue is «(i.e. no customer is lost). Our aim is to find a policy which minimizes the average waiting time of all the customers (which also minimizes the average response time of customers). The problem consists in finding an optimal allocation pattern and to find the optimal ratio («, «). The optimal allocation pattern was found in [1, 3, 2] but no way to compute the optimal ratio is provided there. S 1 Arrivals./D/1 Queue./D/1 Queue S 2 Figure 4: Admission Control in two./d/1 queues The input sequence Û is given under the form of a binary sequence Û such that when Û Òµ the Ò Ø customer is sent in the first queue and when Û Òµ ¼ the Ò Ø customer is sent in the second queue. The average waiting time of all the customers for an input sequence Û is denoted Ï ËË Ûµ. 5.2 Optimal Policies To find an optimal allocation pattern we will use results shown in [2]. Theorem 32. The average waiting time Ï ËË Ûµ is minimized for an lower mechanical input sequence Ñ «ÓÔØ, with some slope «ÓÔØ. Proof. This is a direct consequence of the multi-criteria optimization found in [2]. RR n 3997

22 20 B. Gaujal & E. Hyon Note that the approach proposed in [2] is more general than the system considered here and works for any stationary services times and stationary inter-arrival times. However, it does not provide any mean to compute «ÓÔØ, which is what we are going to do here for deterministic queues. Remark 33. It can be shown that the average waiting time in a deterministic queue under an arrival process of the form Ñ «is the same as the average waiting time when the arrival process is of the form Ñ «. Therefore, Theorem 33 applies with lower mechanical as well as upper mechanical sequences. This remark is essential here since when the input sequence in the first queue is an upper mechanical word of slope «(Ñ «), then the input sequence in the second queue is a lower mechanical word of slope «(Ñ «). Using the considerations of Remark 34, and conditioning on the queue chosen for each customer the relation between Ï ËË Ñ «µ, Ï Ë Ñ «µ and Ï Ë Ñ «µ is 5.3 Optimal Ratio Ï ËË Ñ «µ «Ï Ë Ñ «µ «µ Ï Ë Ñ «µ (29) We will now compute the optimal ratio «ÓÔØ over all possible stable ratios Stability Condition Consider the system of the two./d/1 queues above, the stability condition of such systems is, that is (30) Ë Ë But the stability of the two queues individually is also necessary. Therefore by Equation (5) «Ë and «Ë. Hence for any given number «in Ë Ë the input sequence Ñ «is stable. The interval Special Cases Ë Ë is called the interval of stability in the following. This part considers special degenerated cases where the theory developed in the previous section is not necessary. Lemma 34. In the case Ë Ë, no real optimal policy exists. Proof. The system is never stable since Ë Ë, implying Ë Ë Lemma 35. In the case Ë, an optimal policy consists in sending all the customers in the second queue. In this case Ï ËË Ñ ¼ µ ¼. Proof. When Ë, the service time is smaller than the inter-arrival time. Therefore each time a customer arrives the second queue is empty and we have Ï Ë µ ¼. Since Ï Ë ¼µ ¼ then by Equation (29), Ï ËË Ñ ¼ µ ¼. Lemma 36. If Ë Ë then the round robin policy is an optimal policy. Proof. Lemma (22) implies Ï Ë Ñ µ ¼ and Ï Ë Ñ µ ¼. Therefore a possible optimal ratio is «and Ñ ¼ is an optimal policy. INRIA

23 Optimal Routing Policy in Two Deterministic Queues Case Ë Ë This can be considered as the general case. Characterization of optimal ratio when by: We are interested to find the optimal ratio «ÓÔØ given «ÓÔØ ÑÒ «Ë Ë Ï ËË Ñ «µ Let Ö and Ö denote the jumps of Ï Ë and Ï Ë respectively. Theorem 37. For any real service time Ë and Ë the optimal ratio «ÓÔØ is a jump of Ï Ë or Ï Ë. Hence, it is a rational number and the optimal routing policy is periodic. Proof. By Lemma 31, the function «Ï Ë Ñ «µ is concave for «in Ö Ö and the function «µï Ë Ñ «µ is also concave for «in Ö Ö. Therefore by Equation (29) the function Ï ËË is concave for «in Ö Ö. Where Ö Ö Ö Ë Ë is the set of jumps of Ï ËË. Hence the set of possible minima of Ï ËË is all the jumps Ö or Ë Theorem 41 says that the growing rates of the function Ã Ò exclude the points Ë and Ë of the set of possible minima. or Ë when they are irrational. converges to infinity. This allows us to Therefore the minimum of Ï ËË is a jump and since all the jumps are rational the optimal ratio «ÓÔØ is rational. This means that the optimal routing policy is periodic. Lemma 38. Let Ô be the rational number with the smallest denominator in Ë Ë. Then Ô is the smallest jump of Ï Ë in Ë Ë and the greatest jump of Ï Ë in Ë Ë. Since Ô is the unique common jump of Ï Ë and Ï Ë, we call Ô the double jump. Proof. Let Ö be the greatest jump of Ï Ë Ñ «µ such that Ö Ë. This implies Ë Ö Ë. By Lemma 17 Ö is the rational number with the smallest denominator in Ö Ë, then Ö is the rational number with the smallest denominator in Ë Ë. Let Ö be the greatest jump of Ï Ë «µ such that Ö Ë. This implies Ö is the greatest jump of Ï Ë «µ in Ë Ë and the rational number with the smallest denominator in Ë Ë. Then Ö Ö. The double jump is not always the optimal ratio, as we will see in the example below. Example 39 (Double jump is not optimal). Let Ë and Ë. The mechanical expansion of is and the one of is. We have Ö, Ö, Ö, Ö, Ö Ë and Ö, Ö, Ö, Ö, Ö Ë. The double jump is. The following numerical values have been obtained using exact computations provided by the program presented later (Section 5.4 and 6). Ï Ë Ñ µ ¼ Ï Ë Ñ µ ¼ Ï Ë Ë Ñ µ ¼ Ï Ë Ñ µ ¼ Ï Ë Ñ µ Ï Ë Ë Ñ µ ¼ Since, the optimal ratio is not the double jump. It can also be shown that the optimal ratio in this case is «ÓÔØ. Figure 5 displays the functions Ï Ë Ñ «µ (curve b), Ï Ë Ñ «µ (curve c) and Ï ËË Ñ «µ (curve a), as well as some jumps of those three curves (the double jump 1/5 is marked by the dotted line d, and the optimal jump 2/9 by the dotted line e). The vertical lines f show the interval of stability of the system, Ï ËË Ñ «µ is infinite outside this interval. RR n 3997

24 22 B. Gaujal & E. Hyon a b c d e f Figure 5: Curves of Ï Ë, Ï Ë and Ï ËË. Characterization of optimal ratio when In this part, we will assume that the system is fully loaded. When then the stability interval is reduced to a single point. There is just one ratio «satisfying the stability condition, «Ë. Therefore the upper mechanical word with slope «is optimal. This is the only case where the optimal policy may be aperiodic. 5.4 Algorithm and computational issues We present in Figure 6 an algorithm to compute the optimal ratio «ÓÔØ, when. Find double jump Ô current-jump := Ô Compute the next-jump-right of Ô while Ï ËË Ñ current-jump µ Ï ËË Ñ next-jump-right µ do current-jump:= next-jump-right Compute the next-jump-right of current-jump endwhile Compute the next-jump-left of Ô while Ï ËË Ñ current-jump µ Ï ËË Ñ next-jump-left µ do current-jump:= next-jump-left compute the next-jump-left of current-jump endwhile return current-jump Figure 6: Algorithm computing «ÓÔØ Correctness of the algorithm We will now show that the algorithm is correct and converges to the result in a finite number of steps. For that we have to prove two preliminary lemmas. Lemma 40. The growing rate of the function Ã Ò Proof. We show first by induction that, converges to infinity. Ò Õ Ò Õ (31) INRIA

25 Optimal Routing Policy in Two Deterministic Queues 23 At step one : Ò Õ Ò Õ Ð µ Ð µ Ð Ð µ Ð µ. At step : Ò Õ Ò Õ Ð µõ Ò Õ Ò Ð µõ Ò Ò Õ µ Ò Õ Ò Õ. Hence the growing rate Ã Ò Ã Ò Ò Õ Ò Ò Ã Ò Ã µ Õ Õ µ Õ Ò Ò Ò Ã Ò Ã µ Ò Ã Ò Ã µ Ò Ò Ò µ Ò Ò Ò µ Let us prove now that Ò Ò Ò µ. We use Equations (8) that yield By Equation (12) we obtain Hence Ò Ò Ò µ Ò Ò Ò µ «µ «µ «µ «Ò Ò Ò µ Ð µò Ò Ò µ Ð µò Ò Ò µ Ð Ð Ð Ð µ Ò Ò Ò µ Ò Ò Ò µ Ð µ Ð Ð µ Ð Ð µ Ð «Ð Ð Ð Ð µ Ð µ Ð µ ««Two cases may occur, either the sequence «¼ does not converge to zero or it converges to zero. In the first case the sequence Ò Ò Ò µ does not converges to zero and the series ¼Ò Ò Ò µ goes to infinity. In the other case there exists a number Æ such that Æ «. Since «implies Ð and «µ««, Ò Ò Ò µ Ò Ò Ò µ Ò Ò Ò µ Ò Ò Ò µ Which implies that the sequence is strictly increasing and that the series diverges to infinity. Lemma 41. The function Ã Õ is convex. Proof. The recursive computation of Õ Ñ Ö is given by : We have by Equations (16) and (31) Ã Ã Ò Ò Õ Õ Õ Õ Õ ¼ ¼ Õ Ð Õ Ð µõ Õ Ã Õ Ã Õ Ã Ò Õ Ã Õ Õ Õ Õ µ Ò Õ Ã Õ Ã Õ Ã Õ Ã Õ Ã Õ Ã Õ Õ µ Õ Using Equation (16) leads to Õ Ã Õ Ã Õ Ð µã Ã Õ Ã Õ Ò Ð µ Ð µ Ò Ò µ Õ Ð µã Ã Ð Õ Õ µ Ã Õ Ò Ð µ Ð µ Ò Ò µ Õ Ã Õ Ã Õ Ò Ð µ Ð µ Ò Ò µ RR n 3997

26 24 B. Gaujal & E. Hyon Hence Õ Ã Õ Ã As shown in proof of Lemma 27, This proves the convexity of the function. Õ Ã Õ Ã Õ Ò Ð µ Ð µ Ò Ò µ Ò Ð µ Ð µ Ò Ò µ ¼ (32) Lemma 42. The algorithm is correct and converges in a finite number of steps Proof. Correctness. Since by Lemma 42 the function Ã Õ is convex, then by Equation (29) the function Ï ËË Ñ Ö µ is also convex. Considering that «ÓÔØ Ö shows the correctness of the algorithm. Finiteness. Let be the integers such that Ô Ö Ö Ö. Let us note that Ö Ë Ö. Since Ò Ö Ò Ö Ò, and since by Lemma 41 the growing rate of the function Ï Ë Ñ Ö µ goes to infinity then by Equation (29) Ò the growing rate of the function Ò Ï ËË Ñ Ö Ò µ converges to infinity. Therefore the integer ¼ Ò ¼ such that Ò Ò ¼ Ï ËË Ñ Ö Ò µ Ï ËË Ñ Ö Ò µ ¼ is finite. Using similar arguments for the jumps smaller than Ô shows that the algorithm converges in a finite number of steps. 6 Numerical experiments The algorithm presented above has been implemented in Maple in order to keep exact values for all the rational numbers involved in the computations. This section is dedicated to the presentation of several runs of the program in order to shown how the optimal policy (or equivalently the ratio of the optimal policy) behaves with respect to the parameters of the system, namely Ë, the service time in queue 1, Ë, the service time in queue 2 as well as the inter-arrival time (fixed to one previously, but which can be modified by scaling the time units). In the first series of computations, we fix Ë and Ë. Therefore, the fastest server, Ë is times faster than Ë, with. One could expect that the optimal routing policy sends times more customers in the second queue than in the first queue. This policy has a ratio «, namely «. However, as the experiments in Figure 7 show, the optimal ratio is «ÓÔØ when the inter-arrival time is one. In Figure 7, with Ë and Ë, we let the inter-arrival time vary so that the total load goes from 0 to 1. All the results presented in the figure are exact computations and do not suffer from any numerical errors. When is smaller than (dotted line on the left), many ratios are optimal. For instance, for all «, Ï ËË Ñ «µ ¼. This part is not shown in the figure. The main interest lies within the bounds where the optimal ratio is unique. The optimal ratio «ÓÔØ takes several rational values, ranging from to and ending with the intuitive ratio of the service times,. When, then «is the only point in the stability region. However, for lower values of the load, it is somewhat surprising that the optimal ratio changes with the load and deviates from the ratio of the service times,. The changes of the optimal ratio occur according to several reasons. For example, «ÓÔØ moves away from only when gets out of the stability region (this occurs at ). The same phenomenon occurs when «ÓÔØ jumps away from, which occurs when. On the other hand, the change from 2/9 to 1/5 occurs while 2/9 is still a jump in the stability region. In general, the sudden changes of the optimal ratio when the load increases remain mysterious and deserve further studies. The second set of computations is presented in Figure 8. Here, the inter-arrival time is fixed to one, and the service time in the first queue is fixed to Ë. We let Ë vary from 1 to 51/26 and compute the optimal ratio. INRIA

27 Optimal Routing Policy in Two Deterministic Queues optimal ratios bounds 3/ Figure 7: Optimal ratio when the total load varies Figure 8: Optimal ratio when the service time Ë varies RR n 3997

28 26 B. Gaujal & E. Hyon The Figure shows that «ÓÔØ increases to 1/2 while Ë approaches Ë, which is natural. It also illustrates the high non-linearity of the behavior of the optimal ratio with respect to the parameters of the system. In total, the Maple program computing «ÓÔØ was run hundreds of times, with many rational and irrational parameters. The maximal number of steps before finding the optimal jump never went over 15, which confirms experimentally the intuition that the optimal jump should be in the neighborhood of the double jump in general. However, we believe that one can come up with well chosen parameters for which the optimal jump is arbitrarily far (in number of jumps) from the double jump. Acknowledgement The authors wish to thank D.A. van der Laan for his helpfull comments. References [1] E. Altman, B. Gaujal, and A. Hordijk. Admission control in stochastic event graphs. Technical Report RR 3179, INRIA, June [2] E. Altman, B. Gaujal, and A. Hordijk. Balanced sequences and optimal routing. Technical Report RR 3180, INRIA, January [3] E. Altman, B. Gaujal, and A. Hordijk. Multimodularity, convexity and optimization properties. Mathematics of Operation Research, 25(2), May [4] B. Hajek. Optimal control of two interacting service stations. IEEE Trans. Aut. Cont., 29: , [5] G. Koole, P. Sparaggis, and D. Towsley. Minimizing response times and queue lengths in systems of parrallel queues. Journal of Applied Probability, 36: , [6] M. Lothaire. Algebraic Combinatorics on Words, chapter Sturmian Words. To appear. [7] M. Lothaire. Mots, chapter Tracé de droites, fractions continues et morphismes itérés. Hermes, [8] D. van der Laan. Routing jobs to servers with deterministic distinct service times. Technical report, Leiden University, INRIA

29 Unité de recherche INRIA Lorraine LORIA, Technopôle de Nancy-Brabois - Campus scientifique 615, rue du Jardin Botanique - BP Villers-lès-Nancy Cedex (France) Unité de recherche INRIA Rennes : IRISA, Campus universitaire de Beaulieu Rennes Cedex (France) Unité de recherche INRIA Rhône-Alpes : 655, avenue de l Europe Montbonnot-St-Martin (France) Unité de recherche INRIA Rocquencourt : Domaine de Voluceau - Rocquencourt - BP Le Chesnay Cedex (France) Unité de recherche INRIA Sophia Antipolis : 2004, route des Lucioles - BP Sophia Antipolis Cedex (France) Éditeur INRIA - Domaine de Voluceau - Rocquencourt, BP Le Chesnay Cedex (France) ISSN