- Email: [email protected]
Automatic Workflow Scheduling Tuning for Distributed Processing Systems
Procedia Computer Science Procedia Computer Science 101, 2016, Pages 388 – 397 YSC 2016. 5th International Young Scientist Conference on Computational Science
$XWRPDWLFZRUNIORZVFKHGXOLQJWXQLQJIRUGLVWULEXWHG SURFHVVLQJV\VWHPV Alexander A. Visheratin, Mikhail Melnik and Denis Nasonov ITMO University, Saint Petersburg, Russia [email protected], [email protected], [email protected]
Abstract Modern scientific applications are composed of various methods, techniques and models to solve complicated problems. Such composite applications commonly are represented as workflows. Workflow scheduling is a well-known optimization problem, for which there is a great amount of solutions. Most of the algorithms contain parameters, which affect the result of a method. Thus, for the efficient scheduling it is important to tune parameters of the algorithms. Moreover, performance models, which are used for the estimation of obtained solutions, are crucial parts of workflow scheduling. In this work we present a combined approach for automatic parameters tuning and performance models construction in the background of the WMS lifecycle. Algorithms tuning is provided by hyper-heuristic genetic algorithm, whereas models construction is performed via symbolic regression methods. Developed algorithm was evaluated using CLAVIRE platform and is applicable for any distributed computing systems to optimize the execution of composite applications. Keywords: genetic algorithm, workflow, hyper-heuristic, parameters tuning, performance model.
1 Introduction Complex scientific applications nowadays play a very important role in the development of different scientific domains ± physics, astronomy, biology, etc. ± since they allow users to solve complicated problems by combining various methods and techniques in a single solution. Execution of the composite application (CA) usually involves a large number of smaller applications, which perform some specific tasks and communicate with other parts of the CA through the signals and data transfer. One of the most common ways for representing CA is a workflows concept ± directed acyclic graph, where nodes stand for the computational tasks (parts of the CA) and the edges denote dependencies between them (mostly data connections). For performing execution of such workflows computational capacity of one computer is obviously not enough. High performance computational systems like Grid clusters and cloud environments are used for these purposes. And when speaking about the workflow execution in the distributed systems, very important issue is scheduling of the workflow. Scheduling is a process of mapping workflow tasks onto the computational resources with regard to dependencies between them. It means that task
388
doi:10.1016/j.procs.2016.11.045
Peer-review under responsibility of organizing committee of the scientific committee of the 5th International Young Scientist Conference on Computational Science © 2016 The Authors. Published by Elsevier B.V.
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
ݐଶ , which is dependent of task ݐଵ , cannot be scheduled to be executed before this dependency would be satisfied, e.g. signal or data received from ݐଵ . Workflow scheduling is known to be an NP-hard problem [1], and there are a lot of algorithms, which are designed to perform efficient workflow scheduling. Except of design of the algorithm itself, WKHUH DUH WZR LPSRUWDQW DVSHFWV ZKLFK FDQ JUHDWO\ LQIOXHQFH WKH TXDOLW\ RI WKH DOJRULWKP¶V VROXWLRQ The first is the algorithm parameters. Many algorithms have specific parameters controlling different parts of the execution, like population size, mutation and crossover probabilities for the genetic algorithm. Pertinent parameters can not only increase the quality of the result, but also have positive effect on other characteristics of the algorithm, e.g. decrease convergence time, in case of metaheuristics. The second important aspect is a proper selection of performance models. For the workflow execution in the distributed system there is a plenty of factors to consider ± dependence of the application completion time on the characteristics of the computational resource (CPU cores, memory), the network characteristics (bandwidth, latency), etc. The more of these factors we take into account when creating performance model for the application, the more precise our scheduling algorithm will be. In this paper we present a combined approach, which aims to optimize both scheduling algorithms parameters and performance models used by the algorithms. We implement this approach as a procedure for an automatic tuning of scheduling algorithms and investigate its efficiency in different situations for the previously developed MHGH algorithm [2].
2 Related works In [3] Trelea presents a very detailed analysis on the parameters optimization for the particle swarm optimization algorithm (PSO). In the paper different aspects of the PSO are described and the influence of the algorithm parameters on its convergence behavior are presented. Author makes experimental comparison of the algorithm parameters efficiency between the set of parameters recommended in [4] and the set selected with proposed tuning heuristic after a large number of experiments using five target functions for optimization. There are several important for us results in this paper. The first is an optimal set of parameters for the PSO algorithm, which we can use as a start point for our procedure. The second is the fact that author used heuristic algorithm for searching better parameters set ± despite relatively large execution time, metaheuristic algorithms tend to generate better solutions than heuristics, that is why usage of the GA for these purposes may give better results. $QG WKH WKLUG LV WKH DXWKRU¶V FRQFOXVLRQ WKDW WKH RSWLPDO SDUDPHWHUV¶ VHW VWURQJO\ GHSHQGs on the function being optimized, which makes it very reasonable to adjust algorithms parameters for every problem (in our case workflow) independently. Authors of [5] propose a hyper-heuristic algorithm for tasks scheduling in the cloud environments, which aims to produce results better, than metaheuristics, while not increasing the computation time. The idea of having several metaheuristic algorithms and selection of the algorithm being used on the runtime is very interesting and the results clearly show that this approach generally is more efficient. In the paper Tsai et al. use a fixed set of algorithms, where their parameters do not change over time, which may lead to the potential drawback, because parameters giving good results for one problem may give worse results for the other, and their dynamic adjustment can have significant impact on the resulting solution. Xhafa and Abraham [6] present a comprehensive survey on computational models and heuristic methods for Grid scheduling. They make a thorough look on the different aspects of the tasks execution and scheduling in the Grid environments, such as performance models, optimization criteria and types of scheduling. A wide range of metaheuristic algorithms classes are covered, from local search and population methods to hybrid heuristics and neural networks. Hyper-heuristics presented in the paper are close to [5] and this lack of the algorithms, which can adapt their parameters to the execution environment, gives a room for the solution proposed in our paper.
389
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
In [7] Maheshwari et al. address the problem of workflow scheduling in the multi-site environment, where several high-performance computational clusters are connected with each other through the Internet. Authors proposed a scheduling approach, which tries to tackle the multi-site workflow scheduling using data movement performance modeling. Experimental evaluation shows that the developed algorithm and performance models allow to reduce the total completion time up to 52%. Such significant results show the immense importance of the proper performance models for the scheduling process. Angelova and Pecheva in [8] investigated the tuning of genetic algorithm parameters for the convergence improvement in the problem of fermentation process modeling. Authors used several implementations of the GA and considered the influence of three main parameters ± generation gap, crossover and mutation probabilities ± on the convergence. It was found that the generation gap affects the convergence time the most and well-tuning of this parameter allows to improve the convergence almost to 40%. Authors of [9] propose the hybrid Taguchi-genetic algorithm for tuning the structure and parameters of a feedforward neural network. Global exploration capability of the genetic algorithm combined with the feature of the optimum offspring exploitation of the Taguchi method allows to outperform the GA-based algorithm while reducing the number of connections in the network. Despite the fact that these papers do not directly relate to the workflow scheduling, they demonstrate the universal need for algorithms parameters optimization to achieve prominent results. Herodotou et al. in [10] has developed Starfish ± a self-tuning system for big data analysis based on Hadoop [11]. The system introduces execution tuning on different levels, from the level of jobs to the workloads ± collections of workflows. Authors also propose the method for the runtime optimization of the workload execution through jobs execution profiling. Experimental study shows the efficiency of Starfish in the simultaneous optimization on different levels as well as the ability to optimize different aspects of workload execution ± provisioning, scheduling and data layout. Unfortunately, the described approach is directed only to Hadoop and cannot be directly applied for our case. However, the idea of the workflow execution tracing in order to optimize its execution is very close to our approach. Kumar et al. in [12] propose an integrated framework for workflow execution optimization using variation of user-defined parameters. Authors experimentally show that changing quality-preserving and quality-changing parameters of target workflows leads to significant improvements in its execution time, in some cases more than 50%. The main idea of trading quality of the solution is close to the approach described in [13], but approach of Butakov et al. is more specific for the case of urgent workflows scheduling.
3 Problem statement 3.1 Distributed processing systems In the modern world there is a wide variety of distributed processing systems, which allow users to satisfy almost any needs for executions of their applications. Cloud environments, like Amazon EC2 or Microsoft Azure [14], and Grid platforms, such as Aneka [15] or QCG [16], made it possible for scientists to solve a lot more complicated problems than ever before by describing their complex applications in a distributed manner, which then can be processed by the execution environment. In this paper we focus on the Grid platforms since they provide the ability to create dynamic applications, which execute on the large-scale heterogeneous infrastructure through workflow management system (WMS) [17]. Although Grid systems can have their specific drawbacks, e.g. limited quality of service adherence because of the limited resources, they are widely used in the scientific community, since many research departments have their own computational clusters that can be combined using the WMS. For working with the workflow applications and experimental evaluation of the proposed approach we use the CLAVIRE platform [18]. It provides a solid capability of unifying different resources (PCs,
390
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
clouds, supercomputers, etc.) and describing the composite application in terms of the workflow using the domain specific language EasyFlow. CLAVIRE supports distributed data storage using the developed system DStorage, applications description as a software packages with input/output data and parameters, which then can be used as the parts of the workflow, and composite application creation and execution control through web-based user interface. Since CLAVIRE platform provides a very flexible functionality of working with computational resources and software packages, we were able to design the execution environment for our studies as a Grid platform containing virtual machines deployed on the cluster of our department.
3.2 Workflow scheduling In order to define the problem of the workflow scheduling we have to describe the concept of the workflow and the process of its scheduling on the computational resources of the Grid platform. Scientific workflows are usually represented as a directed acyclic graph in the form ܹ ൌ ሺܶǡ ܦሻ. Here ݐ ܶ אǡ ͳ ݅ ܰ stands for the execution task, which has a performance model and can have specific hardware requirements. ݀ǡ ܦ אǡ ͳ ݅ ܰǡ ͳ ݆ ܰǡ ݅ ് ݆ is a dependency between tasks ݐ and ݐ [19]. The set of computational resources ܴ represents the nodes of the Grid, each of which has execution capabilities in terms of CPU and GPU cores, RAM volume, etc. Workflow scheduling is the process of mapping the set of tasks ܶ on the set of resources ܴ while taking into account the set of dependencies ܦ. Example of the workflow and its scheduling is presented in Figure 1.
Figure 1: (a) Sample workflow containing six tasks; (b) Example of mapping sample workflow on three resources
In Figure 1a a sample workflow containing six tasks is shown. The example of mapping workflow tasks on three resources is presented in Figure 1b. Stroked segments before the start of tasks 4, 5 and 6 represent data transfer from one node to another because of data dependencies between tasks. In our work we make two important considerations. First, we consider only data dependencies between execution steps of the workflow, since they make the largest impact on the workflow scheduling. Second, we assume that only one task can be executed on the node simultaneously. It allows us not to take into account possible resources allocation conflicts between applications (e.g. when several applications, which utilize all available CPU cores run in parallel) and concentrate particularly at the scheduling problem.
3.3 Target composite application For experimental evaluation of the proposed approach we used hydrometeorological composite application (FloodGenerator) designed for analysis of potential flood threats in Saint Petersburg. This application allows users to investigate possible scenarios of the cyclone behavior in the Baltic Sea area and make decisions about the possibility of the flooding well ahead of time. The main idea of the FloodGenerator is the ensemble simulation of synthetic cyclones based on the user-defined parameters with the following generation of water level curves. Workflow of the FloodGenerator is presented in Figure 2.
391
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Figure 2: Schema of the target workflow
7KHILUVWVWHS³3DUDPHWHUVJHQHUDWRU´WDNHVPHDQYDOXHDQGGHYLDWLRQIRUWKHNH\V\QWKHWLFF\FORQH parameters ± center pressure, start point, finish point, etc. ± and generates several sets of parameters using Gaussian distribution. After that for each set of parameters the execution environment creates an instance of the cyclone generator. This application is able to modify HIRLAM forecast files by placing cyclones with specified parameters on the pressure and wind fields thus simulating movement of the cyclone along some path. Detailed theoretical background and key aspects of the cyclone generator can be found in [20]. Resulting HIRLAM files are sent to SWAN wave model that computes random, short-crested wind-generated waves in coastal regions and inland waters [21]. After finishing of SWAN calculation outputs of both SWAN and cyclone generator move to BSM model [22], which outputs the forecast of the water level in the target points of the Baltic Sea. The final step of the ZRUNIORZLV³5HVXOWDQDO\]HU´SDFNDJHZKLFKFROOHFWVUHVXOWVRIDOO ensemble branches and builds a combined plot that can be then analyzed by the expert. Cyclone generator, SWAN and BSM are of the main interest for performance analysis and scheduling LQ WKLV ZRUNIORZ EHFDXVH ³3DUDPHWHUV JHQHUDWRU´DQG³5HVXOWDQDO\]HU´ are quite simple and do not make extensive computations or a lot of data transfer.
4 Proposed solution 4.1 Performance models tuning For the efficient workflow scheduling by metaheuristic algorithms, it is crucial to accurately simulate a computational processes of the workflow execution and estimate the total execution time (makespan) or another features, which can be used as optimization criteria. Consequently, performance models are required to estimate tasks¶UXQWLPHVDQGGDWDWUDQVIHU times between them. In our work we utilize analytical performance models, which are presented in the form of mathematical equations defining the dependence between the execution time of the application and characteristics of computational resources [23]. There are two main approaches for performance model creation ± manual and automatic. Workflow management systems nowadays allow user to manually set up performance models for applications they use (manually define formulas for execution time calculation) in order to make the scheduling more precise. But for doing this user has to understand the way the application works (e.g. single- or multi-thread, dependence on the input data size), and in many cases users do not have such knowledge, because they use a third-party applications and have no access to their internals. In such situation automatic performance model generation can be of a great help. There is a range of methods for creating mathematical equations using the information about
392
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
input and output values ± Lagrange interpolation [24], Vandermonde matrix [25], splines, symbolic regression [26]. In this work, symbolic regression (SR) is proposed for the performance models tuning, because during the preliminary experiments it has shown very good results. The input for the SR is a statistical data about tasks execution on different computational nodes. Using this data SR is able to find a suitable mathematical equation for describing the dependence between the target parameter (e.g. makespan) and parameters of the node (CPU cores, memory) and task (volume of input/output data). This approach combined with the manual performance model specification for the cases of complex relations between parameters allows to accurately estimate the runtime of a certain task on computational resource with a certain configuration.
4.2 Algorithm parameters tuning Workflow scheduling is a well-known optimization problem. A large amount of various algorithms has been developed to solve this problem. These algorithms can be divided into two main classes: heuristic and metaheuristic algorithms. While the most of heuristic algorithms are not-parameterized, the majority of metaheuristics include different sets of parameters. For example, GA operates with crossover and mutation probabilities together with a population size. In a case of PSO, algorithm works with inertia and acceleration coefficients to global and personal bests positions. Algorithms parameters in one or another way affect the efficiency of the algorithms. Since each metaheuristic algorithm can be implemented in various ways, the optimal set of parameters should be tuned for each implementation. For these purposes we propose a hyper-heuristic genetic algorithm (HHGA). The algorithm itself is a standard GA, where each chromosome in the population is a set of parameters for a metaheuristic algorithm under consideration. Mutation and crossover operations modify and combine parameters. Fitness function is calculated by launching the algorithm with parameters obtained from the chromosome and calculating the target optimization parameter (makespan, resources usage cost, etc.). The evolution process is continued for the specified number of iterations (about 300).
4.3 Combined approach The highest efficiency of the algorithm can be achieved only by using the optimal set of DOJRULWKP¶VSDUDPHWHUVZLWKprecise performance models. However, these tunings processes require a lot of computational time and they cannot be applied immediately before a real scheduling process. Thus, our proposed solution assumes that tuning process must be performed in the background. Therefore, algorithms for performance models and algorithms' parameters tuning are launched with a certain period of time to update these models and parameters. In cases, when better solutions are found, they replace corresponding solutions in computing pODWIRUP¶VGDWDEDVH When the conduction of scheduling is required, scheduling algorithm uses the last set of its parameters and performance models for the tasks obtained from the database. Thereby, scheduling algorithms is able to receive the latest configuration and improve its performance.
5 Experiments 5.1 Experimental environment For the experimental study, the data about executions of target CA is obtained from CLAVIRE platform. This data was gathering over two years on various configurations of computing resources and application itself. Log files include such important for us features as runtime, data transfer time, volumes of input and output data for each package of CA.
393
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Obtained statistical data of FloodGenerator CA include executions of packages on computing nodes with 1, 2, 3, 4 cores and with 1, 2, 4, 8 GB of RAM.
5.2 Workflow tasks launches statistics The dependence of package runtime on number of cores or RAM is shown in Figure 3. Each line present executions with certain number of RAM, while X-axis is responsible for number of cores.
Figure 3: Dependence of packages performance on the configuration of resources
It can be clearly seen, that packages Result analyzer (analyzer), BSM and Parameters generator (paramgen) depends only on size of RAM. Cyclone generator (cygen) and SWAN substantially depend on number of CPU cores, but RAM also affects performance of these packages.
5.3 Experimental evaluation Using the data presented in Figure 3, it becomes possible to evaluate performance models of packages depending on computational resources. As it was mentioned in section 4.1, this evaluation is conducted using genetic programming method for symbolic regression. Fitness function of this algorithm is represented as a mean squared error (MSE) for the runtimes of packages. Results of the experiments are presented in Figure 4. From the plots it can be seen that for packages, which were depend basically on RAM, symbolic regression began to converge approximately near 100 iterations of algorithm. For the rest packages (Cyclone generator and SWAN), which depend on number of CPU cores and RAM, symbolic regression shows that algorithm requires greater number of iterations. Moreover, algorithm is not converged completely during 200 iterations, and a better solution may be found by usage of higher number of iterations. During this experimental study, algorithm for tuning of performance models received models with approximate error values: 7%, 3%, 7%, 14% and 8% of mean runtimes for the corresponding packages: Result analyzer, BSM, SWAN, Cyclone generator, Parameters generator.
394
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Figure 4: Error of symbolic regression for performance models
In the next experiment, the hyper-heuristic genetic algorithm (HHGA) is applied to optimize the set of parameters for developed scheduling algorithms: genetic algorithm (GA) and particle swarm optimization (PSO). Developed GA includes two mutation operators and one crossover operator. Therefore, set of parameters for GA includes two mutation probabilities ݉ͳ and ݉ʹ for each mutation operator, crossover probability ܿ and population size . PSO contains inertia coefficient ݓ, two acceleration coefficients ܿͳ and ܿʹ , which present acceleration to personal and global positions respectively, population size and number of neighbors ݊ , which are used to select global best position for each particle at update stage of PSO algorithm. Number of iterations for both algorithms is set to 100 and optimization criteria is total execution time of workflow (makespan). Fitness function is evaluated by simulating the workflow execution process with the usage of performance models, from the previous experiments. Simulated computational environment is consisted of four computing nodes with different configurations of cpu cores and RAM. Target workflow (section 3.3) is chosen as workload with five branches of Cyclone generator, SWAN and BSM packages. The results of experiments are presented in Figure 5. Makespan or fitness function is presented on Y-Axis. The initial set of *$¶V SDUDPHWHUV ݉ͳ, ݉ʹ, ܿ, )was set to (0.6, 0.6, 0.8, 30). HHGA has evaluated (0.32, 0.12, 0.79, 40) as the optimal set of parameters for GA with the 3% profit in comparison to initial parameters. As it turned out, the second mutation operator is proved to be less useful, and its probability is set to 12% during the tuning. 7KHLQLWLDOVHWRI362¶VSDUDPHWHUV ݓǡ ܿͳ, ܿʹ, , ݊) was set to (0.3, 1.2, 0.8, 30, 5), and the result set of parameters, obtained during the tuning is (0.89, 1.12, 1.24, 40, 10). Through the tuning, the efficiency of PSO is increased by 5%.
395
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Figure 5: Performance of hyper-heuristic GA for parameters tuning
It should be noted, that small performance 2% and 5% is explained by comparison of the same algorithms with different sets of parameters. Since initial sets have been chosen adequately for these algorithms, the main objective of HHGA in this case was only to bring these initial sets to optimal states.
6 Conclusion For the efficient workflow scheduling using metaheuristic algorithms, such as GA or PSO, it is crucial to accurately estimate the execution time of tasks DQG WR RSWLPL]H D VHW RI DOJRULWKP¶V parameters. Therefore, in this work, an approach for automated tuning of these performance models DQGDOJRULWKPV¶SDUDPHWHUV was proposed. The automated process works in the background of the Grid platform and updates the database of models and parameters, which is used during the real workflow scheduling. For the performance models construction, symbolic regression was used. Symbolic regression is SHUIRUPHGE\JHQHWLFSURJUDPPLQJDQGXVHVVWDWLVWLFDOGDWDDERXWSDFNDJHV¶H[HFXWLRQV Parameters tuning was performed by hyper-heuristic genetic algorithm. FloodGenerator meteorological application is chosen as a benchmark workflow. The error of obtained performance models varied in range 3-14%. Efficiency of scheduling GA and PSO algorithms is increased by 3% and 5% respectively after their parameters tuning. It results in about 13 minutes makespan improvement for the application with the total runtime about 270 minutes, which can be significant for urgent applications, like flood forecasting and prevention. This paper is financially supported by Ministry of Education and Science of the Russian Federation, Agreement #14.587.21.0024 (18.11.2015) RFMEFI58715X0024.
References [1] [2]
[3] [4] [5]
396
5'XDQ53URGDQDQG7)DKULQJHU³3HUIRUPDQFHDQGFRVWRSWLPL]DWLRQIRUPXOWLSOHODUJHVFDOHJULGZRUNIORZDSSOLFDWLRQV´Proc. 2007 ACM/IEEE Conf. Supercomput. 2007. SC’07., pp. 1±12, 2007. '1DVRQRY1%XWDNRY0%DODNKRQWVHYD..Q\D]NRYDQG$9%RXNKDQRYVN\³+\EULG Evolutionary Workflow Scheduling Algorithm for Dynamic Heterogeneous Distributed &RPSXWDWLRQDO (QYLURQPHQW´ Int. Jt. Conf. SOCO’14-CISIS’14-ICEUTE’14, vol. 239, pp. 83±92, 2014. ,&7UHOHD³7KHSDUWLFOHVZDUPRSWLPL]DWLRQDOJRULWKP &RQYHUJHQFHDQDO\VLVDQGSDUDPHWHU VHOHFWLRQ´Inf. Process. Lett., vol. 85, no. 6, pp. 317±325, 2003. 0&OHUF³7KHVZDUPDQGWKHTXHHQ 7RZDUGV a determininistic and adaptive particle swarm RSWLPL]DWLRQ´Congr. Evol. Comput., vol. 3, pp. 1951±1957, 1999. C.-W. Tsai, W.-C. Huang, M.-H. Chiang, M.-C. Chiang, and C.-6
Automatic workflow scheduling tuning for distributed processing systems
[6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]
[17] [18] [19] [20] [21] [22] [23] [24] [25] [26]
Alexander A. Visheratin et al.
);KDIDDQG$$EUDKDP³&RPSXWDWLRQDOPRGHOVDQGKHXULVWLFPHWKRGVIRU*ULGVFKHGXOLQJ SUREOHPV´Futur. Gener. Comput. Syst., vol. 26, no. 4, pp. 608±621, 2010. K. Maheshwari, E. S. Jung, J. Meng, V. VishwanatK DQG 5 .HWWLPXWKX ³0RGHO-driven PXOWLVLWHZRUNIORZVFKHGXOLQJ´Proc. - IEEE Int. Conf. Clust. Comput. ICCC, 2013. 0$QJHORYDDQG73HQFKHYD³7XQLQJJHQHWLFDOJRULWKPSDUDPHWHUVWRLPSURYHFRQYHUJHQFH WLPH´Int. J. Chem. Eng., vol. 2011, 2011. -77VDL-+&KRXDQG7./LX³7XQLQJWKHVWUXFWXUHDQGSDUDPHWHUVRIDQHXUDOQHWZRUN by using hybrid Taguchi-JHQHWLFDOJRULWKP´IEEE Trans. Neural Networks, vol. 17, no. 1, pp. 69±80, 2006. H. Herodotou, H. Lim, G. Luo, N. Borisov, and /'RQJ³6WDUILVKௗ $6HOI-tuning System for %LJ'DWD$QDO\WLFV´Cidr, vol. 11, pp. 261±272, 2011. T. White, Hadoopࣟ: The Definitive Guide2¶5HLOO\0HGLD,QF V. S. Kumar, P. Sadayappan, G. Mehta, K. Vahi, E. Deelman, V. Ratnakar, J. Kim, Y. Gil, M. +DOO 7 .XUF DQG - 6DOW] ³$Q ,QWHJUDWHG )UDPHZRUN IRU 3DUDPHWHU-based Optimization of 6FLHQWLILF:RUNIORZV´Proc. Int. Symp. High Perform. Distrib. Comput., pp. 177±186, 2009. 1 %XWDNRY ' 1DVRQRY $ 6YLWHQNRY DQG $ 5DGLFH ³4XDOLW\-based Approach to Urgent :RUNIORZV6FKHGXOLQJ´Procedia Comput. Sci., vol. 80, pp. 2074±2085, 2016. & 9HFFKLROD 6 3DQGH\ DQG 5 %X\\D ³+LJK-performance cloud computing: A view of scieQWLILFDSSOLFDWLRQV´I-SPAN 2009 - 10th Int. Symp. Pervasive Syst. Algorithms, Networks, pp. 4±16, 2009. &;LQJFKHQ .1DGLPLQWL -&KDR6 9HQXJRSDODQG5%X\\D ³$QHND 1H[W-generation enterprise grid platform for e-science and e-business appOLFDWLRQV´ Proc. - e-Science 2007, 3rd IEEE Int. Conf. e-Science Grid Comput., pp. 151±159, 2007. ..XURZVNL :'H %DFN:'XELW]N\/*XO\iV*.DPSLV00DPRQVNL *6]HPHV DQG 0 6ZDLQ ³&RPSOH[ V\VWHP VLPXODWLRQV ZLWK 4RV&RV*ULG´ Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 5544 LNCS, no. PART 1, pp. 387±396, 2009. -
397
$XWRPDWLFZRUNIORZVFKHGXOLQJWXQLQJIRUGLVWULEXWHG SURFHVVLQJV\VWHPV Alexander A. Visheratin, Mikhail Melnik and Denis Nasonov ITMO University, Saint Petersburg, Russia [email protected], [email protected], [email protected]
Abstract Modern scientific applications are composed of various methods, techniques and models to solve complicated problems. Such composite applications commonly are represented as workflows. Workflow scheduling is a well-known optimization problem, for which there is a great amount of solutions. Most of the algorithms contain parameters, which affect the result of a method. Thus, for the efficient scheduling it is important to tune parameters of the algorithms. Moreover, performance models, which are used for the estimation of obtained solutions, are crucial parts of workflow scheduling. In this work we present a combined approach for automatic parameters tuning and performance models construction in the background of the WMS lifecycle. Algorithms tuning is provided by hyper-heuristic genetic algorithm, whereas models construction is performed via symbolic regression methods. Developed algorithm was evaluated using CLAVIRE platform and is applicable for any distributed computing systems to optimize the execution of composite applications. Keywords: genetic algorithm, workflow, hyper-heuristic, parameters tuning, performance model.
1 Introduction Complex scientific applications nowadays play a very important role in the development of different scientific domains ± physics, astronomy, biology, etc. ± since they allow users to solve complicated problems by combining various methods and techniques in a single solution. Execution of the composite application (CA) usually involves a large number of smaller applications, which perform some specific tasks and communicate with other parts of the CA through the signals and data transfer. One of the most common ways for representing CA is a workflows concept ± directed acyclic graph, where nodes stand for the computational tasks (parts of the CA) and the edges denote dependencies between them (mostly data connections). For performing execution of such workflows computational capacity of one computer is obviously not enough. High performance computational systems like Grid clusters and cloud environments are used for these purposes. And when speaking about the workflow execution in the distributed systems, very important issue is scheduling of the workflow. Scheduling is a process of mapping workflow tasks onto the computational resources with regard to dependencies between them. It means that task
388
doi:10.1016/j.procs.2016.11.045
Peer-review under responsibility of organizing committee of the scientific committee of the 5th International Young Scientist Conference on Computational Science © 2016 The Authors. Published by Elsevier B.V.
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
ݐଶ , which is dependent of task ݐଵ , cannot be scheduled to be executed before this dependency would be satisfied, e.g. signal or data received from ݐଵ . Workflow scheduling is known to be an NP-hard problem [1], and there are a lot of algorithms, which are designed to perform efficient workflow scheduling. Except of design of the algorithm itself, WKHUH DUH WZR LPSRUWDQW DVSHFWV ZKLFK FDQ JUHDWO\ LQIOXHQFH WKH TXDOLW\ RI WKH DOJRULWKP¶V VROXWLRQ The first is the algorithm parameters. Many algorithms have specific parameters controlling different parts of the execution, like population size, mutation and crossover probabilities for the genetic algorithm. Pertinent parameters can not only increase the quality of the result, but also have positive effect on other characteristics of the algorithm, e.g. decrease convergence time, in case of metaheuristics. The second important aspect is a proper selection of performance models. For the workflow execution in the distributed system there is a plenty of factors to consider ± dependence of the application completion time on the characteristics of the computational resource (CPU cores, memory), the network characteristics (bandwidth, latency), etc. The more of these factors we take into account when creating performance model for the application, the more precise our scheduling algorithm will be. In this paper we present a combined approach, which aims to optimize both scheduling algorithms parameters and performance models used by the algorithms. We implement this approach as a procedure for an automatic tuning of scheduling algorithms and investigate its efficiency in different situations for the previously developed MHGH algorithm [2].
2 Related works In [3] Trelea presents a very detailed analysis on the parameters optimization for the particle swarm optimization algorithm (PSO). In the paper different aspects of the PSO are described and the influence of the algorithm parameters on its convergence behavior are presented. Author makes experimental comparison of the algorithm parameters efficiency between the set of parameters recommended in [4] and the set selected with proposed tuning heuristic after a large number of experiments using five target functions for optimization. There are several important for us results in this paper. The first is an optimal set of parameters for the PSO algorithm, which we can use as a start point for our procedure. The second is the fact that author used heuristic algorithm for searching better parameters set ± despite relatively large execution time, metaheuristic algorithms tend to generate better solutions than heuristics, that is why usage of the GA for these purposes may give better results. $QG WKH WKLUG LV WKH DXWKRU¶V FRQFOXVLRQ WKDW WKH RSWLPDO SDUDPHWHUV¶ VHW VWURQJO\ GHSHQGs on the function being optimized, which makes it very reasonable to adjust algorithms parameters for every problem (in our case workflow) independently. Authors of [5] propose a hyper-heuristic algorithm for tasks scheduling in the cloud environments, which aims to produce results better, than metaheuristics, while not increasing the computation time. The idea of having several metaheuristic algorithms and selection of the algorithm being used on the runtime is very interesting and the results clearly show that this approach generally is more efficient. In the paper Tsai et al. use a fixed set of algorithms, where their parameters do not change over time, which may lead to the potential drawback, because parameters giving good results for one problem may give worse results for the other, and their dynamic adjustment can have significant impact on the resulting solution. Xhafa and Abraham [6] present a comprehensive survey on computational models and heuristic methods for Grid scheduling. They make a thorough look on the different aspects of the tasks execution and scheduling in the Grid environments, such as performance models, optimization criteria and types of scheduling. A wide range of metaheuristic algorithms classes are covered, from local search and population methods to hybrid heuristics and neural networks. Hyper-heuristics presented in the paper are close to [5] and this lack of the algorithms, which can adapt their parameters to the execution environment, gives a room for the solution proposed in our paper.
389
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
In [7] Maheshwari et al. address the problem of workflow scheduling in the multi-site environment, where several high-performance computational clusters are connected with each other through the Internet. Authors proposed a scheduling approach, which tries to tackle the multi-site workflow scheduling using data movement performance modeling. Experimental evaluation shows that the developed algorithm and performance models allow to reduce the total completion time up to 52%. Such significant results show the immense importance of the proper performance models for the scheduling process. Angelova and Pecheva in [8] investigated the tuning of genetic algorithm parameters for the convergence improvement in the problem of fermentation process modeling. Authors used several implementations of the GA and considered the influence of three main parameters ± generation gap, crossover and mutation probabilities ± on the convergence. It was found that the generation gap affects the convergence time the most and well-tuning of this parameter allows to improve the convergence almost to 40%. Authors of [9] propose the hybrid Taguchi-genetic algorithm for tuning the structure and parameters of a feedforward neural network. Global exploration capability of the genetic algorithm combined with the feature of the optimum offspring exploitation of the Taguchi method allows to outperform the GA-based algorithm while reducing the number of connections in the network. Despite the fact that these papers do not directly relate to the workflow scheduling, they demonstrate the universal need for algorithms parameters optimization to achieve prominent results. Herodotou et al. in [10] has developed Starfish ± a self-tuning system for big data analysis based on Hadoop [11]. The system introduces execution tuning on different levels, from the level of jobs to the workloads ± collections of workflows. Authors also propose the method for the runtime optimization of the workload execution through jobs execution profiling. Experimental study shows the efficiency of Starfish in the simultaneous optimization on different levels as well as the ability to optimize different aspects of workload execution ± provisioning, scheduling and data layout. Unfortunately, the described approach is directed only to Hadoop and cannot be directly applied for our case. However, the idea of the workflow execution tracing in order to optimize its execution is very close to our approach. Kumar et al. in [12] propose an integrated framework for workflow execution optimization using variation of user-defined parameters. Authors experimentally show that changing quality-preserving and quality-changing parameters of target workflows leads to significant improvements in its execution time, in some cases more than 50%. The main idea of trading quality of the solution is close to the approach described in [13], but approach of Butakov et al. is more specific for the case of urgent workflows scheduling.
3 Problem statement 3.1 Distributed processing systems In the modern world there is a wide variety of distributed processing systems, which allow users to satisfy almost any needs for executions of their applications. Cloud environments, like Amazon EC2 or Microsoft Azure [14], and Grid platforms, such as Aneka [15] or QCG [16], made it possible for scientists to solve a lot more complicated problems than ever before by describing their complex applications in a distributed manner, which then can be processed by the execution environment. In this paper we focus on the Grid platforms since they provide the ability to create dynamic applications, which execute on the large-scale heterogeneous infrastructure through workflow management system (WMS) [17]. Although Grid systems can have their specific drawbacks, e.g. limited quality of service adherence because of the limited resources, they are widely used in the scientific community, since many research departments have their own computational clusters that can be combined using the WMS. For working with the workflow applications and experimental evaluation of the proposed approach we use the CLAVIRE platform [18]. It provides a solid capability of unifying different resources (PCs,
390
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
clouds, supercomputers, etc.) and describing the composite application in terms of the workflow using the domain specific language EasyFlow. CLAVIRE supports distributed data storage using the developed system DStorage, applications description as a software packages with input/output data and parameters, which then can be used as the parts of the workflow, and composite application creation and execution control through web-based user interface. Since CLAVIRE platform provides a very flexible functionality of working with computational resources and software packages, we were able to design the execution environment for our studies as a Grid platform containing virtual machines deployed on the cluster of our department.
3.2 Workflow scheduling In order to define the problem of the workflow scheduling we have to describe the concept of the workflow and the process of its scheduling on the computational resources of the Grid platform. Scientific workflows are usually represented as a directed acyclic graph in the form ܹ ൌ ሺܶǡ ܦሻ. Here ݐ ܶ אǡ ͳ ݅ ܰ stands for the execution task, which has a performance model and can have specific hardware requirements. ݀ǡ ܦ אǡ ͳ ݅ ܰǡ ͳ ݆ ܰǡ ݅ ് ݆ is a dependency between tasks ݐ and ݐ [19]. The set of computational resources ܴ represents the nodes of the Grid, each of which has execution capabilities in terms of CPU and GPU cores, RAM volume, etc. Workflow scheduling is the process of mapping the set of tasks ܶ on the set of resources ܴ while taking into account the set of dependencies ܦ. Example of the workflow and its scheduling is presented in Figure 1.
Figure 1: (a) Sample workflow containing six tasks; (b) Example of mapping sample workflow on three resources
In Figure 1a a sample workflow containing six tasks is shown. The example of mapping workflow tasks on three resources is presented in Figure 1b. Stroked segments before the start of tasks 4, 5 and 6 represent data transfer from one node to another because of data dependencies between tasks. In our work we make two important considerations. First, we consider only data dependencies between execution steps of the workflow, since they make the largest impact on the workflow scheduling. Second, we assume that only one task can be executed on the node simultaneously. It allows us not to take into account possible resources allocation conflicts between applications (e.g. when several applications, which utilize all available CPU cores run in parallel) and concentrate particularly at the scheduling problem.
3.3 Target composite application For experimental evaluation of the proposed approach we used hydrometeorological composite application (FloodGenerator) designed for analysis of potential flood threats in Saint Petersburg. This application allows users to investigate possible scenarios of the cyclone behavior in the Baltic Sea area and make decisions about the possibility of the flooding well ahead of time. The main idea of the FloodGenerator is the ensemble simulation of synthetic cyclones based on the user-defined parameters with the following generation of water level curves. Workflow of the FloodGenerator is presented in Figure 2.
391
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Figure 2: Schema of the target workflow
7KHILUVWVWHS³3DUDPHWHUVJHQHUDWRU´WDNHVPHDQYDOXHDQGGHYLDWLRQIRUWKHNH\V\QWKHWLFF\FORQH parameters ± center pressure, start point, finish point, etc. ± and generates several sets of parameters using Gaussian distribution. After that for each set of parameters the execution environment creates an instance of the cyclone generator. This application is able to modify HIRLAM forecast files by placing cyclones with specified parameters on the pressure and wind fields thus simulating movement of the cyclone along some path. Detailed theoretical background and key aspects of the cyclone generator can be found in [20]. Resulting HIRLAM files are sent to SWAN wave model that computes random, short-crested wind-generated waves in coastal regions and inland waters [21]. After finishing of SWAN calculation outputs of both SWAN and cyclone generator move to BSM model [22], which outputs the forecast of the water level in the target points of the Baltic Sea. The final step of the ZRUNIORZLV³5HVXOWDQDO\]HU´SDFNDJHZKLFKFROOHFWVUHVXOWVRIDOO ensemble branches and builds a combined plot that can be then analyzed by the expert. Cyclone generator, SWAN and BSM are of the main interest for performance analysis and scheduling LQ WKLV ZRUNIORZ EHFDXVH ³3DUDPHWHUV JHQHUDWRU´DQG³5HVXOWDQDO\]HU´ are quite simple and do not make extensive computations or a lot of data transfer.
4 Proposed solution 4.1 Performance models tuning For the efficient workflow scheduling by metaheuristic algorithms, it is crucial to accurately simulate a computational processes of the workflow execution and estimate the total execution time (makespan) or another features, which can be used as optimization criteria. Consequently, performance models are required to estimate tasks¶UXQWLPHVDQGGDWDWUDQVIHU times between them. In our work we utilize analytical performance models, which are presented in the form of mathematical equations defining the dependence between the execution time of the application and characteristics of computational resources [23]. There are two main approaches for performance model creation ± manual and automatic. Workflow management systems nowadays allow user to manually set up performance models for applications they use (manually define formulas for execution time calculation) in order to make the scheduling more precise. But for doing this user has to understand the way the application works (e.g. single- or multi-thread, dependence on the input data size), and in many cases users do not have such knowledge, because they use a third-party applications and have no access to their internals. In such situation automatic performance model generation can be of a great help. There is a range of methods for creating mathematical equations using the information about
392
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
input and output values ± Lagrange interpolation [24], Vandermonde matrix [25], splines, symbolic regression [26]. In this work, symbolic regression (SR) is proposed for the performance models tuning, because during the preliminary experiments it has shown very good results. The input for the SR is a statistical data about tasks execution on different computational nodes. Using this data SR is able to find a suitable mathematical equation for describing the dependence between the target parameter (e.g. makespan) and parameters of the node (CPU cores, memory) and task (volume of input/output data). This approach combined with the manual performance model specification for the cases of complex relations between parameters allows to accurately estimate the runtime of a certain task on computational resource with a certain configuration.
4.2 Algorithm parameters tuning Workflow scheduling is a well-known optimization problem. A large amount of various algorithms has been developed to solve this problem. These algorithms can be divided into two main classes: heuristic and metaheuristic algorithms. While the most of heuristic algorithms are not-parameterized, the majority of metaheuristics include different sets of parameters. For example, GA operates with crossover and mutation probabilities together with a population size. In a case of PSO, algorithm works with inertia and acceleration coefficients to global and personal bests positions. Algorithms parameters in one or another way affect the efficiency of the algorithms. Since each metaheuristic algorithm can be implemented in various ways, the optimal set of parameters should be tuned for each implementation. For these purposes we propose a hyper-heuristic genetic algorithm (HHGA). The algorithm itself is a standard GA, where each chromosome in the population is a set of parameters for a metaheuristic algorithm under consideration. Mutation and crossover operations modify and combine parameters. Fitness function is calculated by launching the algorithm with parameters obtained from the chromosome and calculating the target optimization parameter (makespan, resources usage cost, etc.). The evolution process is continued for the specified number of iterations (about 300).
4.3 Combined approach The highest efficiency of the algorithm can be achieved only by using the optimal set of DOJRULWKP¶VSDUDPHWHUVZLWKprecise performance models. However, these tunings processes require a lot of computational time and they cannot be applied immediately before a real scheduling process. Thus, our proposed solution assumes that tuning process must be performed in the background. Therefore, algorithms for performance models and algorithms' parameters tuning are launched with a certain period of time to update these models and parameters. In cases, when better solutions are found, they replace corresponding solutions in computing pODWIRUP¶VGDWDEDVH When the conduction of scheduling is required, scheduling algorithm uses the last set of its parameters and performance models for the tasks obtained from the database. Thereby, scheduling algorithms is able to receive the latest configuration and improve its performance.
5 Experiments 5.1 Experimental environment For the experimental study, the data about executions of target CA is obtained from CLAVIRE platform. This data was gathering over two years on various configurations of computing resources and application itself. Log files include such important for us features as runtime, data transfer time, volumes of input and output data for each package of CA.
393
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Obtained statistical data of FloodGenerator CA include executions of packages on computing nodes with 1, 2, 3, 4 cores and with 1, 2, 4, 8 GB of RAM.
5.2 Workflow tasks launches statistics The dependence of package runtime on number of cores or RAM is shown in Figure 3. Each line present executions with certain number of RAM, while X-axis is responsible for number of cores.
Figure 3: Dependence of packages performance on the configuration of resources
It can be clearly seen, that packages Result analyzer (analyzer), BSM and Parameters generator (paramgen) depends only on size of RAM. Cyclone generator (cygen) and SWAN substantially depend on number of CPU cores, but RAM also affects performance of these packages.
5.3 Experimental evaluation Using the data presented in Figure 3, it becomes possible to evaluate performance models of packages depending on computational resources. As it was mentioned in section 4.1, this evaluation is conducted using genetic programming method for symbolic regression. Fitness function of this algorithm is represented as a mean squared error (MSE) for the runtimes of packages. Results of the experiments are presented in Figure 4. From the plots it can be seen that for packages, which were depend basically on RAM, symbolic regression began to converge approximately near 100 iterations of algorithm. For the rest packages (Cyclone generator and SWAN), which depend on number of CPU cores and RAM, symbolic regression shows that algorithm requires greater number of iterations. Moreover, algorithm is not converged completely during 200 iterations, and a better solution may be found by usage of higher number of iterations. During this experimental study, algorithm for tuning of performance models received models with approximate error values: 7%, 3%, 7%, 14% and 8% of mean runtimes for the corresponding packages: Result analyzer, BSM, SWAN, Cyclone generator, Parameters generator.
394
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Figure 4: Error of symbolic regression for performance models
In the next experiment, the hyper-heuristic genetic algorithm (HHGA) is applied to optimize the set of parameters for developed scheduling algorithms: genetic algorithm (GA) and particle swarm optimization (PSO). Developed GA includes two mutation operators and one crossover operator. Therefore, set of parameters for GA includes two mutation probabilities ݉ͳ and ݉ʹ for each mutation operator, crossover probability ܿ and population size . PSO contains inertia coefficient ݓ, two acceleration coefficients ܿͳ and ܿʹ , which present acceleration to personal and global positions respectively, population size and number of neighbors ݊ , which are used to select global best position for each particle at update stage of PSO algorithm. Number of iterations for both algorithms is set to 100 and optimization criteria is total execution time of workflow (makespan). Fitness function is evaluated by simulating the workflow execution process with the usage of performance models, from the previous experiments. Simulated computational environment is consisted of four computing nodes with different configurations of cpu cores and RAM. Target workflow (section 3.3) is chosen as workload with five branches of Cyclone generator, SWAN and BSM packages. The results of experiments are presented in Figure 5. Makespan or fitness function is presented on Y-Axis. The initial set of *$¶V SDUDPHWHUV ݉ͳ, ݉ʹ, ܿ, )was set to (0.6, 0.6, 0.8, 30). HHGA has evaluated (0.32, 0.12, 0.79, 40) as the optimal set of parameters for GA with the 3% profit in comparison to initial parameters. As it turned out, the second mutation operator is proved to be less useful, and its probability is set to 12% during the tuning. 7KHLQLWLDOVHWRI362¶VSDUDPHWHUV ݓǡ ܿͳ, ܿʹ, , ݊) was set to (0.3, 1.2, 0.8, 30, 5), and the result set of parameters, obtained during the tuning is (0.89, 1.12, 1.24, 40, 10). Through the tuning, the efficiency of PSO is increased by 5%.
395
Automatic workflow scheduling tuning for distributed processing systems
Alexander A. Visheratin et al.
Figure 5: Performance of hyper-heuristic GA for parameters tuning
It should be noted, that small performance 2% and 5% is explained by comparison of the same algorithms with different sets of parameters. Since initial sets have been chosen adequately for these algorithms, the main objective of HHGA in this case was only to bring these initial sets to optimal states.
6 Conclusion For the efficient workflow scheduling using metaheuristic algorithms, such as GA or PSO, it is crucial to accurately estimate the execution time of tasks DQG WR RSWLPL]H D VHW RI DOJRULWKP¶V parameters. Therefore, in this work, an approach for automated tuning of these performance models DQGDOJRULWKPV¶SDUDPHWHUV was proposed. The automated process works in the background of the Grid platform and updates the database of models and parameters, which is used during the real workflow scheduling. For the performance models construction, symbolic regression was used. Symbolic regression is SHUIRUPHGE\JHQHWLFSURJUDPPLQJDQGXVHVVWDWLVWLFDOGDWDDERXWSDFNDJHV¶H[HFXWLRQV Parameters tuning was performed by hyper-heuristic genetic algorithm. FloodGenerator meteorological application is chosen as a benchmark workflow. The error of obtained performance models varied in range 3-14%. Efficiency of scheduling GA and PSO algorithms is increased by 3% and 5% respectively after their parameters tuning. It results in about 13 minutes makespan improvement for the application with the total runtime about 270 minutes, which can be significant for urgent applications, like flood forecasting and prevention. This paper is financially supported by Ministry of Education and Science of the Russian Federation, Agreement #14.587.21.0024 (18.11.2015) RFMEFI58715X0024.
References [1] [2]
[3] [4] [5]
396
5'XDQ53URGDQDQG7)DKULQJHU³3HUIRUPDQFHDQGFRVWRSWLPL]DWLRQIRUPXOWLSOHODUJHVFDOHJULGZRUNIORZDSSOLFDWLRQV´Proc. 2007 ACM/IEEE Conf. Supercomput. 2007. SC’07., pp. 1±12, 2007. '1DVRQRY1%XWDNRY0%DODNKRQWVHYD..Q\D]NRYDQG$9%RXNKDQRYVN\³+\EULG Evolutionary Workflow Scheduling Algorithm for Dynamic Heterogeneous Distributed &RPSXWDWLRQDO (QYLURQPHQW´ Int. Jt. Conf. SOCO’14-CISIS’14-ICEUTE’14, vol. 239, pp. 83±92, 2014. ,&7UHOHD³7KHSDUWLFOHVZDUPRSWLPL]DWLRQDOJRULWKP &RQYHUJHQFHDQDO\VLVDQGSDUDPHWHU VHOHFWLRQ´Inf. Process. Lett., vol. 85, no. 6, pp. 317±325, 2003. 0&OHUF³7KHVZDUPDQGWKHTXHHQ 7RZDUGV a determininistic and adaptive particle swarm RSWLPL]DWLRQ´Congr. Evol. Comput., vol. 3, pp. 1951±1957, 1999. C.-W. Tsai, W.-C. Huang, M.-H. Chiang, M.-C. Chiang, and C.-6
Automatic workflow scheduling tuning for distributed processing systems
[6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]
[17] [18] [19] [20] [21] [22] [23] [24] [25] [26]
Alexander A. Visheratin et al.
);KDIDDQG$$EUDKDP³&RPSXWDWLRQDOPRGHOVDQGKHXULVWLFPHWKRGVIRU*ULGVFKHGXOLQJ SUREOHPV´Futur. Gener. Comput. Syst., vol. 26, no. 4, pp. 608±621, 2010. K. Maheshwari, E. S. Jung, J. Meng, V. VishwanatK DQG 5 .HWWLPXWKX ³0RGHO-driven PXOWLVLWHZRUNIORZVFKHGXOLQJ´Proc. - IEEE Int. Conf. Clust. Comput. ICCC, 2013. 0$QJHORYDDQG73HQFKHYD³7XQLQJJHQHWLFDOJRULWKPSDUDPHWHUVWRLPSURYHFRQYHUJHQFH WLPH´Int. J. Chem. Eng., vol. 2011, 2011. -77VDL-+&KRXDQG7./LX³7XQLQJWKHVWUXFWXUHDQGSDUDPHWHUVRIDQHXUDOQHWZRUN by using hybrid Taguchi-JHQHWLFDOJRULWKP´IEEE Trans. Neural Networks, vol. 17, no. 1, pp. 69±80, 2006. H. Herodotou, H. Lim, G. Luo, N. Borisov, and /'RQJ³6WDUILVKௗ $6HOI-tuning System for %LJ'DWD$QDO\WLFV´Cidr, vol. 11, pp. 261±272, 2011. T. White, Hadoopࣟ: The Definitive Guide2¶5HLOO\0HGLD,QF V. S. Kumar, P. Sadayappan, G. Mehta, K. Vahi, E. Deelman, V. Ratnakar, J. Kim, Y. Gil, M. +DOO 7 .XUF DQG - 6DOW] ³$Q ,QWHJUDWHG )UDPHZRUN IRU 3DUDPHWHU-based Optimization of 6FLHQWLILF:RUNIORZV´Proc. Int. Symp. High Perform. Distrib. Comput., pp. 177±186, 2009. 1 %XWDNRY ' 1DVRQRY $ 6YLWHQNRY DQG $ 5DGLFH ³4XDOLW\-based Approach to Urgent :RUNIORZV6FKHGXOLQJ´Procedia Comput. Sci., vol. 80, pp. 2074±2085, 2016. & 9HFFKLROD 6 3DQGH\ DQG 5 %X\\D ³+LJK-performance cloud computing: A view of scieQWLILFDSSOLFDWLRQV´I-SPAN 2009 - 10th Int. Symp. Pervasive Syst. Algorithms, Networks, pp. 4±16, 2009. &;LQJFKHQ .1DGLPLQWL -&KDR6 9HQXJRSDODQG5%X\\D ³$QHND 1H[W-generation enterprise grid platform for e-science and e-business appOLFDWLRQV´ Proc. - e-Science 2007, 3rd IEEE Int. Conf. e-Science Grid Comput., pp. 151±159, 2007. ..XURZVNL :'H %DFN:'XELW]N\/*XO\iV*.DPSLV00DPRQVNL *6]HPHV DQG 0 6ZDLQ ³&RPSOH[ V\VWHP VLPXODWLRQV ZLWK 4RV&RV*ULG´ Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 5544 LNCS, no. PART 1, pp. 387±396, 2009. -
397