Positioning traffic in NoSQL database systems by the use of particle swarm algorithm Marcin Woźniak Institute of Mathematics Silesian University of Technology Kaszubska 23, 44-100 Gliwice, Poland Email: Marcin.Wozniak@polsl.pl Abstract—In this paper, application of particle swarm al- (using methods like [21] or [20]) server responses to the gorithm in positioning and optimization of traffic in NoSQL requests, but this goes according to the income queue. Earlier database is discussed. Sample system is modeled with independent requests must be served first and others according to arrival 2-order hyper exponential input stream of packets and exponen- time. The problem is to position this system for most efficient tial service time distribution. Optimization is solved using particle operation (we shall define optimal service, vacation and in- swarm algorithm for various scenarios of operation. come parameters). In this paper NoSQL database system will be positioned for most efficient service and lowest possible cost I. I NTRODUCTION of work by the use of particle swarm optimization algorithm In modern computer science, artificial intelligence (AI) (PSO). as well as evolutionary computing (EC) is one of most important fields, widely applied in various tasks. There are II. A PPLIED MASS SERVICE MODEL many applications of AI in sciences and industry. The power of such approaches lies within the dedicated mechanisms is For NoSQL database systems various methods of modeling used to simulate sophisticated phenomenon. On the other and simulation can be applied. Mainly we try to analyze hand some techniques of EC were proven very efficient for the model, which describes operation. Operation model is searching optimal solutions, easy to implement and precise. defined for applied queueing system (QS). Service description Let us give same examples. AI applied to create learning of NoSQL database, where dedicated QS is applied to optimize sets are discussed in [1], [2]. Some aspects of positioning operation cost defines Tservice , Tincome and Tvacation , which computing network models by the use of EC are presented describe average time of service, average income time and in [3], [4] and [5]. Moreover, AI is used for the optimization average vacation time (backup, conservation and etc.), respec- of industry processes [6], [7], [8]. AI is also applied for tively. All these are independent random variables, where the systems which require dedicated solutions [9], [10], [11], as symbols in time t are: well as for agents oriented programming and object oriented • τ1 — the first busy period starting at t = 0; refactoring techniques [12], [13], [14], [15]. In NoSQL systems • δ1 — the first idle time (first vacation time and first we use dedicated solutions to increase performance. These standby time); applications are especially designed for given purposes, please • h(τ1 ) — the number of packets served during τ1 ; see [16]. We must build applications to serve clients requests, • X(t) — the number of packets in the system at t. search in database, maintain service and more. All these aspects demand special mechanisms like dedicated sorting In this paper is discussed simulation and positioning of NoSQL and indexing algorithms and queueing systems to administrate database traffic modeled with dedicated QS, where we define incoming requests. Dedicated sorting algorithms help to or- only one request arrival and response departure. ganize large data sets as fast as possible. Some examples of dedicated sorting algorithms are discussed in [17], [18], [19] A. Analytical results and [20]. While in [18] and [20] is presented dedicated version of quick sort, where implemented modifications enabled faster Modeling traffic in NoSQL systems is non-trivial problem. sorting. In [17] is discussed dedicated merge sort, which even Classical cost structure is considered in [25]. While in [26], more efficient version is presented in [21]. Moreover extended [27], [28] are presented most important aspects of positioning research on these situations are discussed in [22] and [23], and cost optimization. Various queueing models for applied where research on efficient methods of indexing and sorting type of the server are investigated in [29], [30], [31], [32], large NoSQL systems are presented. Here will be discussed [33], [34], [35]. Please see also [36] and [37] for a review of another important aspect of optimal service in large NoSQL important results on modeling and positioning. systems - traffic simulation and positioning. In this paper are applied results of the research on similar Traffic in the network and therefore efficient service can objects, see [38] and [39] for joint transform of first busy increase Quality of Service (QoS) [24]. We can simulate period, first idle time and number of packets completely served the network traffic, where NoSQL database server is serving during first busy period in GI/G/1-type systems. More on various clients. Clients send requests and server collects them generally distributed service times and infinite buffers can be to proceed actions. After processing knowledge in database found in [40] and [41]. All these research results are helpful to model and position QS of different type as discussed in [42] then for model of traffic finally we have: [43], [4] and [44]. Where in [42] or [43] was given an idea to apply evolutionary computation (EC) in QS simulation and ∂ En τ1 = − Bn (s, 0, 1) , (6) positioning. An extension of the research for sophisticated QS ∂s s=0 were published in [4]. And finally main analytical model with similarly we have: detailed description and assessments for traffic in the system was given in [44]. Let us see the model of QS for NoSQL ∂ En δ1 = − Bn (0, %, 1) . (7) database traffic. ∂% %=0 To model NoSQL server operation was used a finite- buffer H2 /M/1/N -type QS, similar to server traffic modeling III. A PPLIED PARTICLE S WARM O PTIMIZATION ALGORITHM functions discussed in [45] and [46]. Let it be here presented only a brief description, just to help in understanding NoSQL Particle swarm optimization algorithm (PSO) has been positioning and simulation problem (for details please see shown in [47]. In the initial form PSO was modeling behaviors [44]). Incoming requests describes 2-order distribution func- that can be observed in young birds or fish, which in the tion: cluster behave in a very specific way. Thanks to the ease of implementation and adaptation to different tasks PSO algo- F (t) = p1 1 − e−λ1 t + p2 1 − e−λ2 t , t > 0, (1)   rithm has become one of the most commonly used algorithms where λi > 0 for i = 1, 2 and p1 , p2 ≥ 0. Inter-arrival times in it’s original or modified versions. In [48] is presented are mixed of two exponential distributions with parameters adaptivity of this methods to different initial conditions of λ1 and λ2 , which are being “chosen” with probabilities p1 positioned object. While in [49], [50] and [51] many possible and p2 . In the system, there are (N − 1) places in queue aspects of application of various EC methods in engineering and one for packet in the service. System starts working at optimization are discussed. In [52] PSO application in relay t = 0 with at least one packet present. After busy period the times is presented. Finally discussion on theoretical aspects of server begins vacation which is modeled with 2-order hyper convergence and stability can be found in [53], [54] and [55]. exponential distribution function: Let us discuss behavior of typical swarm. V (t) = q1 1 − e−α1 t + q2 1 − e−α2 t , t > 0. (2)   In the swarm similar operations are performed by many individuals of the same species. In action, individuals com- Interpretation of parameters αi , i = 1, 2 and q1 , q2 is similar municate with each other in a manner characteristic for the to that for λi , i = 1, 2 and p1 and p2 . If at the end of vacation species. Communication helps to exchange information and as there is no packet present in the system, the server is on a result the whole swarm is moving in a certain direction or standby and waits for first arrival to start service process. If behaves like one big organism. PSO algorithm uses the insights there is at least one packet waiting for service in the buffer that emerge from the observation of swarms of fish or insects at the end of vacation, the service process starts immediately that are looking for food or a safe place. This process can and new busy period begins. be described in mathematical model. If we accept the goal of optimizing criterion function of the object, we can talk about Functions F (·) and V (·) help to simulate operation of optimizing algorithm. the examined NoSQL system, where inter-arrival times and vacation are defined in (1) and (2). In the research PSO is used PSO algorithm searches the space of test solutions by to find optimal set of parameters λi , pi , µ and αi . To describe matching trajectories of individuals (particles) in a quasi- minimal amount of resources to perform all operations rn (c1 ) stochastic way. A particular individual is a vector and it’s is defined: movement is the result of stochastic and deterministic com- ponents of movement model. Stochastic component corre- Qn (c1 ) r(τ1 )En τ1 + r(δ1 )En δ1 sponds to random walk. In contrast, deterministic component rn (c1 ) = = , (3) En (c1 ) En τ1 + En δ1 of the movement model is distance between particles, or other feature, which is modeled in mathematical equation. In where the symbols are: r(τ1 )-fixed unit operation costs during subsequent periods individual particles move in looking for busy period τ1 , r(δ1 )-fixed unit operation costs during idle the global optimum, where because of stochastic component time δ1 , En τ1 -means of busy period τ1 and En δ1 -idle time δ1 this movement also has a random character. This combination on condition that system starts with n packets present. In (3) gives ability to efficiently search the test area of the simulated are used means of busy period and vacation (idle) time. The or positioned object. If during motion the particle is on a explicit formula with detailed information and description for new position, which is characterized by better properties of conditional joint characteristic functions of τ1 , δ1 and h(τ1 ) is the optimum, for this position it updates the knowledge. In presented in [4] and [44]. Here let us briefly discuss modeling further exploration particle accepts found value as the optimum of applied QS. General equation to calculate this values is: and starts searching in relation to this value. In each iteration Bn (s, %, z) = E{e−sτ1 −%δ1 z h(τ1 ) | X(0) = n}, 2 ≤ n ≤ N, of the algorithm, the particles can communicate with each (4) other and share information about the sought optimum. If we where s ≥ 0, % ≥ 0 and |z| ≤ 1, n ≥ 1. Details on this consider that all particles in the swarm want to reach the sought equation are discussed in [17], [4] and [44], where using it we optimum criterion function for the positioned object, in the end can define, components of (3) total cos of work: we can take as the optimum best of all-values. In this way, entire swarm is communicating between it’s individuals while En e−sτ1 = E{e−sτ1 | X(0) = n} = Bn (s, 0, 1), (5) looking for the global optimum of the criterion function. A. PSO model Algorithm 1 Basic PSO applied to position NoSQL database system traffic Actions taken while searching for the optimum criterion 1: Define all coefficients: α–optimum value memory factor, function of the object are written as mathematical equations. β–optimum position memory factor, generation– number The model of particle swarm movement keeps the communica- of iterations in the algorithm, particles–number of parti- tion between particles based on a deterministic factor, but also cles in the swarm, introduces randomness of the movements. To build the model 2: Dedicated criterion function: lowest cost of NoSQL system of the swarm behavior in the solution space of the object are operation (3), used the following assumptions: 3: Create at random initial population, 4: t:=0, • Points in the search space are seen as potential solu- 5: while t ≤ generations do tions to moving particle swarm. 6: Move particles according to (9) and (8), 7: Sort particles according to the value of criterion func- • Each particle is seeking for optimum, which is deter- tion, mined by it’s position in the space. 8: Evaluate population and take best ratio of them to next • At the end of PSO iteration, the particles interact with generation, other particles and change information. 9: Rest of particles take at random, 10: Next generation: t + +, • As a result of communication global optimum is 11: end while selected, relative to which all particles are continuing 12: Best particles from the last generation are potential their search. optimum. • Number of moving particles is determined. IV. R ESEARCH RESULTS In the model, we mean a particle moving in a virtual way. We only model the choice of the optimum to which particle has Research results help to predict possible response time and moved. Selected points in the study area are compared, and optimize service cost rn (c1 ) considered in different variants: among them is chosen the global optimum, see also [56] for under-load, critical load and overload. PSO simulations were details on convergence and stability of PSO. performed for r(τ1 ) = 0.5 and r(δ1 ) = 0.5. It means that modeled NoSQL database system uses 0.5 energy unit PSO algorithm for each particle takes the form of xti whose each vacation and work period. For other system types these i components correspond to dimensions of the test space. Each values may be changed in (3), what makes presented model particle is moving at the speed vit appropriate for the swarm flexible and easily applicable. All presented research results in a particular PSO iteration. These values vary in subsequent are averaged values of 100 PSO samplings for 20 particles iterations of the algorithm. Speed of movement of the particles in 80 iterations with α = 0.4 and β = 0.4. In each iteration is described by the formula: best ratio = 90%, what means that 72 best particles were moved to next generation and 8 were taken at random. This vit+1 = vit + α · 1 · [g∗t − f (xti )] + β · 2 · [xt∗ − xti ], (8) helped to search entire object space for optimum values, where: where the symbols are: vit –speed of i particle in t iteration, • Average service time: Tservice = µ1 , α–optimum value memory factor, β–optimum position mem- • Average time between packages income into the sys- ory factor, 1 , 2 ∈ [0, 1]–random values, g∗t –optimum for t tem: Tincome = λp11 + λp22 , iteration, xt∗ –optimum position for t iteration, f (xti )– fitness function value for i particle in t iteration, xti –position of i • Average vacation time: Tvacation = αq11 + αq22 , particle in t iteration. • Examined system size: N = buffer size +1. Equation repositioning particle swarm movement in each Scenario 1. iteration of the PSO algorithm is defined using formula: PSO was performed to find set of parameters for lowest cost of work. In Table I are optimum values for all parameters that xt+1 i = xti + (−1)K · vit , (9) affect NoSQL server work. PSO positioned NoSQL system where the symbols are: xti –position of i particle in t iteration, TABLE I. O PTIMAL PARAMETERS µ, λi , αi , pi , qi FOR i = 1, 2 AND vit –speed of i particle in t iteration, K–random factor to change LOWEST VALUE OF (3). motion direction. The initial coordinates of the particle swarm λ1 λ2 α1 α2 p1 p2 q1 q2 position and their speed we take at random. However, it is 2.9 2.3 1.43 0.32 1.78 1.3 6.1 3.5 possible also to apply some boundary criteria that will allow µ 0.6 rn (c1 ) 0.34 additional control of the swarm. Tservice Tincome Tvacation [sec] 1.67 1.18 15.20 These two equations allow to change position of each particle and therefore search entire space for the optimum of to operate at minimum costs, if the service and vacation are the modeled object. Let us now see possible implementation results from Table I. PSO was also arranged to position the of PSO, which is presented in Algorithm 1. system in various scenarios. Scenario 2. average on-line shop or customer service). Calculated values NoSQL Tservice = 2[sec], what means that request service of Tservice and Tincome gave positioning for lowest cost of takes about 2[sec]. Research results are shown in Table II. work. If the system works with calculated parameters QoS is still very high, but also cost of service is possibly lowest, TABLE II. O PTIMAL PARAMETERS µ, λi , αi , pi , qi FOR i = 1, 2 AND what means better profit for the owner. In the article, have LOWEST VALUE OF (3). been examined newly proposed methods for QS simulation λ1 λ2 α1 α2 p1 p2 q1 q2 and positioning (see also [4] and [44]). EC methods like PSO 2.13 3.15 0.94 0.78 79.70 0.89 2.30 12.10 are excellent for simulation or positioning of different objects. µ 0.5 rn (c1 ) 0.37 PSO method helps to simulate complicated objects and because Tservice Tincome Tvacation [sec] 2.08 37.70 17.96 of the free design, calculations are easy to perform. The experi- ments confirmed PSO efficiency in simulation and positioning the system in various scenarios representing common traffic situations. Scenario 3. NoSQL Tservice = 0.5[sec]. This situation represents NoSQL business service with heavy traffic and very efficient server R EFERENCES machine. Research results with system positioning are shown [1] G. Capizzi, C. Napoli, and L. Paternò, “An innovative hybrid neuro- in Table III. wavelet method for reconstruction of missing data in astronomical TABLE III. O PTIMAL PARAMETERS µ, λi , αi , pi , qi FOR i = 1, 2 AND photometric surveys,” in Artificial Intelligence and Soft Computing. LOWEST VALUE OF (3). Springer Berlin Heidelberg, 2012, pp. 21–29. [2] C. Napoli, F. Bonanno, and G. Capizzi, “An hybrid neuro-wavelet λ1 λ2 α1 α2 p1 p2 q1 q2 approach for long-term prediction of solar wind,” IAU Symposium, no. 44.3 22.1 112.9 1.4 1.91 1 57.60 14.3 274, pp. 247–249, 2010. µ 2.00 rn (c1 ) 0.29 [3] C. Napoli, G. Pappalardo, E. Tramontana, Z. Marszałek, D. Połap, Tservice Tincome Tvacation and M. Woźniak, “Simplified firefly algorithm for 2d image key-points [sec] 0.5 0.09 10.72 search,” in IEEE Symposium Series on Computational Intelligence. IEEE, 2014. Using PSO it is also possible to position the system for [4] M. Woźniak, W. M. Kempa, M. Gabryel, R. K. Nowicki, and Z. Shao, Tincome . This will correspond to peculiar incoming traffic. “On applying evolutionary computation methods to optimization of vacation cycle costs in finite-buffer queue,” Lecture Notes in Artificial Scenario 4. Intelligence - ICAISC’2014, vol. 8467 (PART I), pp. 480–491, 2014. NoSQL Tincome was given as 2[sec], what means that requests [5] C. Napoli, F. Bonanno, and G. Capizzi, “Exploiting solar wind time are incoming to the server once in every 2 seconds. Research series correlation with magnetospheric response by using an hybrid results are shown in Table IV. neuro-wavelet approach,” Proceedings of the International Astronomical Union, vol. 6, no. S274, pp. 156–158, 2010. TABLE IV. O PTIMAL PARAMETERS µ, λi , αi , pi , qi FOR i = 1, 2 AND [6] G. Capizzi, F. Bonanno, and C. Napoli, “Hybrid neural networks LOWEST VALUE OF (3). architectures for soc and voltage prediction of new generation batter- ies storage,” in Clean Electrical Power (ICCEP), 2011 International λ1 λ2 α1 α2 p1 p2 q1 q2 3.7 4.5 1.3 1.4 6.2 2.5 13.1 7.2 Conference on. IEEE, 2011, pp. 341–344. µ 0.4 rn (c1 ) 0.43 [7] F. Bonanno, G. Capizzi, A. Gagliano, and C. Napoli, “Optimal man- Tservice Tincome Tvacation agement of various renewable energy sources by a new forecasting [sec] 2.5 2.23 15.22 method,” in Power Electronics, Electrical Drives, Automation and Motion (SPEEDAM), 2012 International Symposium on. IEEE, 2012, pp. 934–940. Scenario 5. [8] F. Bonanno, G. Capizzi, G. L. Sciuto, C. Napoli, G. Pappalardo, and NoSQL Tincome was given as 0.5[sec], what means that E. Tramontana, “A cascade neural network architecture investigating requests are incoming to the server twice in every second. surface plasmon polaritons propagation for thin metals in openmp,” This situation is describing an extensively used database, Lecture Notes in Artificial Intelligence - ICAISC’2014, vol. 8468, PART I, pp. 22–33, 2014. like these of business purpose. Research results are shown in [9] G. Capizzi, F. Bonanno, and C. Napoli, “Recurrent neural network- Table V. based control strategy for battery energy storage in generation systems TABLE V. O PTIMAL PARAMETERS µ, λi , αi , pi , qi FOR i = 1, 2 AND with intermittent renewable energy sources,” in Clean Electrical Power LOWEST VALUE OF (3). (ICCEP), 2011 International Conference on. IEEE, 2011, pp. 336–340. [10] F. Bonanno, G. Capizzi, and C. Napoli, “Some remarks on the appli- λ1 λ2 α1 α2 p1 p2 q1 q2 cation of rnn and prnn for the charge-discharge simulation of advanced 27.1 27.2 0.8 0.4 16.1 1.1 5.6 4.3 lithium-ions battery energy storage,” in Power Electronics, Electrical µ 0.7 rn (c1 ) 0.33 Drives, Automation and Motion (SPEEDAM), 2012 International Sym- Tservice Tincome Tvacation posium on. IEEE, 2012, pp. 941–945. [sec] 1.43 0.63 17.75 [11] G. Capizzi, F. Bonanno, and C. Napoli, “A new approach for lead-acid batteries modeling by local cosine,” in Power Electronics Electrical Drives Automation and Motion (SPEEDAM), 2010 International Sym- posium on. IEEE, 2010, pp. 1074–1079. V. C ONCLUSIONS [12] C. Napoli, G. Pappalardo, and E. Tramontana, “Using modularity metrics to assist move method refactoring of large systems,” in Sev- Positioned model was simulated in situations with prede- enth International Conference on Complex, Intelligent, and Software fined time of service or time of income. Given times reflect Intensive Systems - CISIS 2013, July 2013, pp. 529–534. situations when traffic is heavy and system must serve many [13] R. Giunta, G. Pappalardo, and E. Tramontana, “AODP: refactoring code requests (like business machines) or common traffic (like to provide advanced aspect-oriented modularization of design patterns,” in Proceedings of Symposium on Applied Computing (SAC). ACM, [33] E. Tramontana, “Automatically characterising components with con- 2012. cerns and reducing tangling,” in Proceedings of QUORS workshop at [14] E. Tramontana, “Detecting extra relationships for design patterns roles,” Compsac. IEEE, 2013. in Proceedings of AsianPlop, March 2014. [34] Z. Niu and Y. Takahashi, “A finite-capacity queue with exhaus- [15] G. Pappalardo and E. Tramontana, “Suggesting extract class refactoring tive vacation/close-down/setup times and markovian arrival processes,” opportunities by measuring strength of method interactions,” in Pro- Queueing Systems, vol. 1, no. 31, pp. 1–23, 1999. ceedings of Asia Pacific Software Engineering Conference (APSEC). [35] Z. Niu, T. Shu, and Y. Takahashi, “A vacation queue with setup and IEEE, December 2013. close-down times and batch markovian arrival processes,” Performance Evaluation Journal, vol. 3, no. 54, pp. 225–248, 2003. [16] Z. Marszałek and M. Woźniak, “On possible organizing nosql database systems,” International Journal of Information Science and Intelligent [36] H. Takagi, Queueing Analysis, vol. 1: Vacation and Priority Systems, System, vol. 2, no. 2, pp. 51–59, 2013. vol. 2. Finite Systems. Amsterdam: North-Holland, 1993. [17] M. Woźniak, Z. Marszałek, M. Gabryel, and R. K. Nowicki, “Modified [37] N. Tian and Z. G. Zhang, Vacation queueing models. Theory and merge sort algorithm for large scale data sets,” Lecture Notes in applications. Berlin, Heidelberg: Springer - Verlag, 2006. Artificial Intelligence - ICAISC’2013, vol. 7895 (PART II), pp. 612– [38] W. M. Kempa, “Gi/g/1/ batch arrival queuing system with a single 622, 2013. exponential vacation,” Mathematical Methods of Operations Research, [18] ——, “On quick sort algorithm performance for large data sets,” in vol. 1, no. 69, pp. 81–97, 2009. Looking into the Future of Creativity and Decision Support Systems, [39] ——, “Characteristics of vacation cycle in the batch arrival queuing A. M. J. Skulimowski, Ed. Cracow, Poland: Progress & Business system with single vacations and exhaustive service,” International Publishers, 2013, pp. 647–656. Journal of Applied Mathematics, vol. 4, no. 23, pp. 747–758, 2010. [19] ——, “Triple heap sort algorithm for large data sets,” in Looking [40] ——, “Some new results for departure process in the m/g/1 queuing into the Future of Creativity and Decision Support Systems, A. M. J. system with a single vacation and exhaustive service,” Stochastic Skulimowski, Ed. Cracow, Poland: Progress & Business Publishers, Analysis and Applications, vol. 1, no. 28, pp. 26–43, 2009. 2013, pp. 657–665. [41] ——, “On departure process in the batch arrival queue with single [20] ——, “Preprocessing large data sets by the use of quick sort algorithm,” vacation and setup time,” Annales UMCS Informatica, vol. 1, no. 10, Advances in Intelligent Systems and Computing - KICSS’2013, vol. pp. 93–102, 2010. accepted–in press, 2014. [42] M. Gabryel, R. K. Nowicki, M. Woźniak, and W. M. Kempa, “Genetic [21] Z. Marszałek, D. Połap, and M. Woźniak, “On preprocessing large cost optimization of the gi/m/1/n finite-buffer queue with a single data sets by the use of triple merge sort algorithm,” in Proceedings of vacation policy,” Lecture Notes in Artificial Intelligence - ICAISC’2013, International Conference on Advances in Information Processing and vol. 7895 (PART II), pp. 12–23, 2013. Communication Technologies - IPCT’2014. Santa Barbara, California, [43] M. Woźniak, “On applying cuckoo search algorithm to positioning USA: The IRED, Seek Digital Library, 2014, pp. 65–72. gi/m/1/n finite-buffer queue with a single vacation policy,” in Pro- [22] M. Woźniak and Z. Marszałek, Selected Algorithms for Sorting Large ceedings of the 12th Mexican International Conference on Artificial Data Sets. Gliwice, Poland: Silesian University of Technology Press, Intelligence - MICAI’2013. IEEE, 2013, pp. 59–64. 2013. [44] M. Woźniak, W. M. Kempa, M. Gabryel, and R. K. Nowicki, “A [23] ——, Extended Algorithms for Sorting Large Data Sets. Gliwice, finite-buffer queue with single vacation policy - analytical study with Poland: Silesian University of Technology Press, 2014. evolutionary positioning,” International Journal of Applied Mathematics and Computer Science, vol. 24, no. 4, pp. accepted–in press, 2014. [24] C. Napoli, G. Pappalardo, and E. Tramontana, “A hybrid neurowavelet predictor for qos control and stability,” in Proceedings of AI*IA, ser. [45] D. Hongwei, Z. Dongfeng, and Z. Yifan, “Performance analysis of LNCS, vol. 8249. Springer, 2013, pp. 527–538. wireless sensor networks of serial transmission mode with vacation on fire prevention,” ICCET’10 IEEE CPS, pp. 153–155, 2010. [25] J. Teghem, “Control of the service process in a queueing system,” European Journal of Operations Research, vol. 1, no. 23, pp. 141–158, [46] V. Mancuso and S. Alouf, “Analysis of power saving with continuous 1986. connectivity,” Computer Networks, vol. 56, no. 10, pp. 2481–2493, 2012. [26] O. Kella, “Optimal control of the vacation scheme in an m/g/1 queue,” [47] J. Kennedy and R. C. Eberhard, “Particie swarm optimization,” in IEEE Operations Research Journal, vol. 4, no. 38, pp. 724–728, 1990. International Conference on Neural Networks. Piscataway, NJ: IEEE, [27] R. Lillo, “Optimal operating policy for an m/g/1 exhaustive server- 1995, pp. 1942–1948. vacation model,” Methodology and Computing in Applied Probability, [48] M. Hu, T. Wu, and J. Weir, “An adaptive particle swarm optimization vol. 2, no. 2, pp. 153–167, 2000. with multiple adaptive methods,” IEEE Transactions on Evolutionary [28] F. Bonanno, G. Capizzi, G. L. Sciuto, C. Napoli, G. Pappalardo, and Computation, vol. 17, no. 5, pp. 705–720, 2013. E. Tramontana, “A novel cloud-distributed toolbox for optimal energy [49] S. Koziel and X. Yang, Computational Optimization, Methods and dispatch management from renewables in igss by using wrnn predictors Algorithms. Berlin, Heidelberg: Springer, 2011. and gpu parallel solutions,” in Power Electronics, Electrical Drives, Automation and Motion (SPEEDAM), 2014 International Symposium [50] M. Gabryel, M. Woźniak, and R. K. Nowicki, “Creating learning sets on. IEEE, 2014, pp. 1077–1084. for control systems using an evolutionary method,” Lecture Notes in Computer Science - ICAISC’2012, vol. 7269, pp. 206–213, 2012. [29] U. C. Gupta, A. D. Banik, and S. Pathak, “Complete analysis of map/g/1/n queue with single (multiple) vacation(s) under limited service [51] X. Yang, Engineering Optimisation: An Introduction with Metaheuristic discipline,” Journal of Applied Mathematics and Stochastic Analysis, Applications. USA: John Wiley & Sons, 2010. no. 3, pp. 353–373, 2005. [52] J. Bansal and K. Deep, “Optimisation of directional overcurrent relay times by particle swarm optimisation,” in SIS’2008 Proceedings. IEEE, [30] U. C. Gupta and K. Sikdar, “Computing queue length distributions in 2008, pp. 1–7. map/g/1/n queue under single and multiple vacation,” Applied Mathe- matics and Computation, vol. 2, no. 174, pp. 1498–1525, 2006. [53] E. Baonabeau, M. Dorigo, and G. Theraulaz, Swarm Intelligence: From Natural to Artificial Systems. Oxford University Press, 1999. [31] C. Napoli, G. Papplardo, and E. Tramontana, “Improving files availabil- ity for bittorrent using a diffusion model,” in IEEE 23nd International [54] V. Gazi and K. Passino, Swarm stability and optimization. Berlin, Workshop on Enabling Technologies: Infrastructure for Collaborative Heidelberg: Springer, 2011. Enterprises - WETICE 2014, June 2014, pp. 191–196. [55] X. Yang, Z. Cui, R. Xiao, A. Gandomi, and M. Karamanoglu, Swarm [32] F. Banno, D. Marletta, G. Pappalardo, and E. Tramontana, “Tack- Intelligence and Bio-inspired Computation: Theory and Applications. ling consistency issues for runtime updating distributed systems,” in OXFORD, United Kingdom: Elsevier, 2013. Proceedings of International Symposium on Parallel & Distributed [56] M. Clerc and J. Kennedy, “The particle swarmexplosion, stability and Processing, Workshops and Phd Forum (IPDPSW). IEEE, 2010, pp. convergence in a multidimensional complex space,” IEEE Transactions 1–8. on Evolutionary Computation, vol. 6, no. 1, pp. 58–73, 2002.