Search

Systems deployed on elastic infrastructures deal with resource variations by adapting themselves, which may cause errors, or even degrade their performance. Therefore, we must test the Cloud-Based Systems(CBSs) throughout elasticity, which faces problematics. First, executing elasticity tests may require the lead of CBS throughout a specific elastic behavior, i. e.,sequence of resource changes, which depends on an accurate workload generation. Second, CBS adaptations occur at a precise moment, such as after a resource scale out, which requires to test them during a specific CBS states. Third, testers must re-execute elasticity tests in a deterministic manner to debug and fix the CBS. Fourth, implementing elasticity tests is complex and laborious given the wide possibility of parameters, and the peculiarity of cloud computing. Finally, specific combinations of parameters may cause the system issues, where random tests may miss such combinations, while a test set that covers all thecombinations may be large and impractical to execute. In this thesis, we tackle all these five problematics by proposing several approaches: 1) an approach to drive the CBS throughout preset elastic behaviors, 2) an approach to synchronize tests according to the CBS states, 3) an approach to enable reproducing elasticity testing, 4) a Domain Specific Language (DSL)-basedapproach to abstract the elasticity testing implementation, and 5) an approach to generate small sets of tests that reveal relevant elasticity-related issues.; Les systèmes déployés dans être testés pendant l’élasticité, ce qui entraîne plusieurs problématiques. D’abord, l’exécution d’un test pendant l’élasticité peut exiger de conduire le CBS dans une succession de comportements élastiques spécifiques, càd., une séquence d’ajout/retrait de ressources, qui nécessite des variations précises de la charge des requêtes envoyées au cloud. Seconde, certaines adaptations du CBS ne sont réalisées qu’à un moment précis, par exemple après un ajout de ressources et, par conséquent, leurs tests doivent être synchronisés avec des états spécifiques du CBS. Troisième, les testeurs doivent rejouer les tests pendant l’élasticité de manière déterministe afin de déboguer et corriger le CBS. Quatrième, la création des tests pendant l’élasticité est complexe et laborieuse dû au large nombre de paramètres, et à la particularité du cloud computing. Enfin, seulement quelques combinaisons de paramètres peuvent causer des problèmes au CBS, que les cas de test créés au hasard peuvent manquer, alors qu’un jeu de tests couvrant toutes les combinaisons possibles serait trop grand et impossible à exécuter. Dans cette thèse, nous abordons toutes ces problématiques en proposant plusieurs approches :1) une approche qui conduit les CBSs dans une suite de comportements élastiques prédéfinis, 2) une approche qui synchronise l’exécution des tests selon les états du CBS, 3) une approche qui permette la reproduction des tests pendant l’élasticité, 4) un langage spécifique à ce domaine (DSL, selon l’acronyme anglais) qui résume la mise en œuvre des tests pendant l’élasticité, 5) une approche qui génère des petits ensembles de tests pendant l’élasticité tout en révélant des problèmes liés à l’élasticité.

Systems deployed on elastic infrastructures deal with resource variations by adapting themselves, which may cause errors, or even degrade their performance. Therefore, we must test the Cloud-Based Systems(CBSs) throughout elasticity, which faces problematics. First, executing elasticity tests may require the lead of CBS throughout a specific elastic behavior, i. e.,sequence of resource changes, which depends on an accurate workload generation. Second, CBS adaptations occur at a precise moment, such as after a resource scale out, which requires to test them during a specific CBS states. Third, testers must re-execute elasticity tests in a deterministic manner to debug and fix the CBS. Fourth, implementing elasticity tests is complex and laborious given the wide possibility of parameters, and the peculiarity of cloud computing. Finally, specific combinations of parameters may cause the system issues, where random tests may miss such combinations, while a test set that covers all thecombinations may be large and impractical to execute. In this thesis, we tackle all these five problematics by proposing several approaches: 1) an approach to drive the CBS throughout preset elastic behaviors, 2) an approach to synchronize tests according to the CBS states, 3) an approach to enable reproducing elasticity testing, 4) a Domain Specific Language (DSL)-basedapproach to abstract the elasticity testing implementation, and 5) an approach to generate small sets of tests that reveal relevant elasticity-related issues.; Les systèmes déployés dans être testés pendant l’élasticité, ce qui entraîne plusieurs problématiques. D’abord, l’exécution d’un test pendant l’élasticité peut exiger de conduire le CBS dans une succession de comportements élastiques spécifiques, càd., une séquence d’ajout/retrait de ressources, qui nécessite des variations précises de la charge des requêtes envoyées au cloud. Seconde, certaines adaptations du CBS ne sont réalisées qu’à un moment précis, par exemple après un ajout de ressources et, par conséquent, leurs tests doivent être synchronisés avec des états spécifiques du CBS. Troisième, les testeurs doivent rejouer les tests pendant l’élasticité de manière déterministe afin de déboguer et corriger le CBS. Quatrième, la création des tests pendant l’élasticité est complexe et laborieuse dû au large nombre de paramètres, et à la particularité du cloud computing. Enfin, seulement quelques combinaisons de paramètres peuvent causer des problèmes au CBS, que les cas de test créés au hasard peuvent manquer, alors qu’un jeu de tests couvrant toutes les combinaisons possibles serait trop grand et impossible à exécuter. Dans cette thèse, nous abordons toutes ces problématiques en proposant plusieurs approches :1) une approche qui conduit les CBSs dans une suite de comportements élastiques prédéfinis, 2) une approche qui synchronise l’exécution des tests selon les états du CBS, 3) une approche qui permette la reproduction des tests pendant l’élasticité, 4) un langage spécifique à ce domaine (DSL, selon l’acronyme anglais) qui résume la mise en œuvre des tests pendant l’élasticité, 5) une approche qui génère des petits ensembles de tests pendant l’élasticité tout en révélant des problèmes liés à l’élasticité.

Web applications are often exposed to unpredictable work-loads, which makes computing resource management di-cult. The resource may be overused when the workload is high and underused when the workload is low. A solution to deal with unpredictable workloads is to migrate Web applications to cloud computing infrastructures, where the resource is varied according to demand, i. e., elasticity. With elasticity, all the resource variations happen during the Web application runtime. To deal with this, the Web application , and its service layers must behave in an elastic manner , which comprises adaptation tasks. These tasks may introduce functional and non-functional errors into the Web application. To nd these errors, we must test the Web application when the adaptation tasks are performed, during the resource variations. Some tests may require a specic sequence of resource variations, which are dicult to be achieved without controllability. Therefore, in this paper, we propose an approach that controls the required resource variations. We validated our approach by conducting several experiments on Amazon EC2 cloud infrastructures. In these experiments, we successfully lead the Web application through the required resource variations.

International audience; Applications that are exposed to fluctuating workloads must ensure elasticity, that is the ability to scale up and down rapidly to meet the demand. Cloud infrastructures provide adaptation tasks, which allow applications to automatically scale up and down straightforwardly. These adaptation tasks drive the system to new states, which may expose implementation errors and therefore must be tested. In this paper, we focus on testing elastic applications during different elasticity-related states. This test is difficult since the elasticity states are not directly controlled by the tester. To execute the test at different elasticity-related states, we propose a controlled testing procedure. This procedure consists in controlling the leading of applications through the required elasticity states, while they are tested accordingly. To test the application, we identify the elasticity states at real time, and in parallel, we execute the state-related tests. We validate our testing procedure by performing experiments on Amazon EC2. In the experiments, we successfully lead an application through the required elasticity states, and identify non-functional errors.

International audience; Web applications are often exposed to unpredictable workloads, which makes computing resource management difficult. The resource may be overused when the workload is high and underused when the workload is low. A solution to deal with unpredictable workloads is to migrate Web applications to cloud computing infrastructures, where the resource is varied according to demand, \ie, elasticity. With elasticity, all the resource variations happen during the Web application runtime. To deal with this, the Web application, and its service layers must behave in an elastic manner, which comprises adaptation tasks. These tasks may introduce functional and non-functional errors into the Web application. To find these errors, we must test the Web application when the adaptation tasks are performed, during the resource variations. Some tests may require a specific sequence of resource variations, which are difficult to be achieved without controllability. Therefore, in this paper, we propose an approach that controls the required resource variations. We validated our approach by conducting several experiments on Amazon EC2 cloud infrastructures. In these experiments, we successfully lead the Web application through the required resource variations.

National audience; MapReduce (MR) is the most popular solution to build applications for large-scale data processing. These applications are often deployed on large clusters of commodity machines, where failures happen constantly due to bugs, hardware problems, and outages. Testing MR-based systems is hard, since it is needed a great eort of test harness to execute distributed test cases upon failures. In this paper, we present a novel testing solution to tackle this issue called HadoopTest. This solution is based on a scalable harness approach, where distributed tester components are hung around each map and reduce worker (i.e., node). Testers are allowed to stimulate each worker to inject failures on them, monitor their behavior, and validate testing results. HadoopTest was used to test two applications bundled into Hadoop, the Apache open source MapReduce implementation. Our initial implementation demonstrates promising results, with HadoopTest coordinating test cases across distributed MapReduce workers, and finding bugs.

Orientador : Prof. Dr. Eduardo Cunha de Almeida Dissertação (mestrado) - Universidade Federal do Paraná, Setor de Ciências Exatas, Programa de Pós-Graduação em Informática. Defesa: Curitiba, 08/12/2011 Bibliografia: fls. 51-45 Resumo: O MapReduce (MR) e uma das soluções mais populares na área de processamento de dados em grande escala. Os sistemas baseados em MapReduce frequentemente são implantados sobre agrupamentos de computadores, onde falhas acontecem constantemente, devido a defeitos de software, problemas de hardware e interrupções. Testar sistemas baseados em MapReduce é difícil, uma vez que é necessário um grande esforço do controlador de testes para executar casos de teste distribuídos em ambientes com a presença de falhas. Neste trabalho, apresentamos uma nova solução de testes para resolver isso, que foi chamada de HadoopTest. Esta solução baseia-se em uma abordagem de controle escalável, onde um coordenador gerência diversos testadores distribuídos, que controlam os componentes do MR. Os testadores podem simular falhas sobre os componentes do MR e monitorar suas execuções. O HadoopTest foi utilizado para testar duas aplicações distribuídas juntamente com o Hadoop (i.e., a implementação MapReduce de código aberto mantida pela fundação Apache). Nossos experimentos apresentaram resultados promissores, sendo que o HadoopTest conseguiu coordenar casos de teste distribuídos, injetar falhas nos componentes do MR e encontrar alguns defeitos de software que foram propositalmente inseridos. Abstract: MapReduce (MR) is one of the most popular solution on large-scale data processing area. The MR-based systems are often deployed over clusters of computers, where failures happen constantly due to bugs, hardware problems, and outages. Testing MR-based systems is hard, since it is needed a great eort of test controller to execute distributed test cases upon failures. In this work, we present a novel testing solution to tackle this issue called HadoopTest. This solution is based on a scalable control approach, where a coordinator manages many distributed testers which control the MR components. Testers are allowed to simulate failures on MR components and monitor their behavior. HadoopTest was used to test two applications bundled into Hadoop (i.e., a open source MapReduce implementation mantained by Apache Foundation). On our experiments HadoopTest was able to coordinate distributed test cases, inject faults on MR components and nd some bugs which were purposely inserted.