Tomasz Buchert Lucas Nussbaum INRIA, LORIA, Universit e de - PowerPoint PPT Presentation

Leveraging business workflows in distributed systems research for the orchestration of reproducible and scalable experiments Tomasz Buchert Lucas Nussbaum INRIA, LORIA, Universit´ e de Lorraine, CNRS T. Buchert, L. Nussbaum Business workflows in distributed systems research 1 / 21

What is a distributed system? A distributed system is one in which the failure of a computer you didn’t even know existed can render your own computer unusable. Leslie Lamport T. Buchert, L. Nussbaum Business workflows in distributed systems research 2 / 21

What is a distributed system? (2) Examples of well-known distributed systems: DNS, BitTorrent, Gmail. T. Buchert, L. Nussbaum Business workflows in distributed systems research 3 / 21

What is a distributed system? (3) Distributed systems are: complex, erroneous, difficult to control, nondeterministic. Moreover, they: span many networks, span many geographical locations, consist of thousands of machines. T. Buchert, L. Nussbaum Business workflows in distributed systems research 4 / 21

Research in distributed systems Can we make better BitTorrent, Gmail, etc.? Solution: experimentation . The “classical way” is difficult, but other solutions exist: simulation, benchmarking, emulation. Real program Modeled program Real platform Standard approach Benchmarking Modeled platform Emulation Simulation T. Buchert, L. Nussbaum Business workflows in distributed systems research 5 / 21

Existing tools Many solutions exist (hardware and software based): Grid’5000, Emulab, PlanetLab, OMF, Expo, Plush, g5k-campaign, ... among the others. How can one evaluate or grade them? T. Buchert, L. Nussbaum Business workflows in distributed systems research 6 / 21

Our final goal To improve the research, the experimentation framework has to: improve descriptiveness of the experiments , handle unexpected , but inevitable errors, ensure scalability of experiments, ensure reproducibility of the results. In the end, we want to improve experimentation on testbeds like Grid’5000 and PlanetLab. T. Buchert, L. Nussbaum Business workflows in distributed systems research 7 / 21

Agenda Introduction Issues with distributed systems research Approach – a framework to control experiments Our contributions: Identify requirements for an experimental framework 1 A new approach based on Business Process Management 2 Conclusions T. Buchert, L. Nussbaum Business workflows in distributed systems research 8 / 21

Contribution 1 Let’s define properties an experimentation framework must have . That way we can evaluate existing and future solutions. We distinguished 12 features in 3 categories. T. Buchert, L. Nussbaum Business workflows in distributed systems research 9 / 21

Design To improve the process of designing the experiment we need: descriptiveness, modularity, reusability, maintainability, support for common patterns. T. Buchert, L. Nussbaum Business workflows in distributed systems research 10 / 21

Execution Required features of experiment execution are: snapshotting, error handling, integration with lower-level tools, human interaction. T. Buchert, L. Nussbaum Business workflows in distributed systems research 11 / 21

Monitoring Finally, to monitor the experiment, we need: monitoring, instrumentation, data analysis. T. Buchert, L. Nussbaum Business workflows in distributed systems research 12 / 21

By the way... The partitioning of the properties is not arbitrary. The three categories: design, execution, monitoring, map to phases in Business Process Management Lifecycle . T. Buchert, L. Nussbaum Business workflows in distributed systems research 13 / 21

Contribution 2 None of existing solutions covers all our requirements. Can we provide them all at once ? Our approach is based on Business Process Management (BPM). T. Buchert, L. Nussbaum Business workflows in distributed systems research 14 / 21

Business Process Management Traditionally, Business Process Management is about: understanding how an organization works, modeling its activities in a language of workflows , executing processes and monitoring their progress, identifying ways to improve activities in an organization, redesigning processes to make them: cheaper , faster , less defective . T. Buchert, L. Nussbaum Business workflows in distributed systems research 15 / 21

Toyota Production System Toyota Corporation is (2010): the largest producer of cars world-wide, the most profitable car-making company, well known for its socio-technical system of practices. The fundamental objective is to eliminate waste or muda ( 無 駄 ): overproduction, waiting, transportation, processing, stock at hand, movement and defective products. T. Buchert, L. Nussbaum Business workflows in distributed systems research 16 / 21

BPM and experimentation There is an analogy between production and experimentation : a car Ð Ñ a finished experiment, an assembly line Ð Ñ execution of an experiment, parts of the assebly line Ð Ñ parts of an experiment, eliminated waste (muda) Ð Ñ efficiency and reliability. T. Buchert, L. Nussbaum Business workflows in distributed systems research 17 / 21

Example of a workflow Create Update system Configure Create VO & CA Evaluate configuration and add gLite sites and set up VOMS TORQUE files repositories III a III ‘ III ‘ Configure worker nodes III ‘ + + Start services Prepare configuration T. Buchert, L. Nussbaum Business workflows in distributed systems research 18 / 21

How BPM fulfills the requirements? Create Update system Configure configuration and add gLite Create VO & CA Evaluate sites files repositories and set up VOMS TORQUE III a III ‘ III ‘ Configure worker nodes III ‘ + + Start services Prepare configuration T. Buchert, L. Nussbaum Business workflows in distributed systems research 19 / 21

How BPM fulfills the requirements? Create Update system Configure configuration and add gLite Create VO & CA Evaluate sites files repositories and set up VOMS TORQUE III a III ‘ III ‘ Configure worker nodes III ‘ + + Start services Prepare configuration Design Descriptiveness Modularity Reusability Maintainability Support for common patterns T. Buchert, L. Nussbaum Business workflows in distributed systems research 19 / 21

How BPM fulfills the requirements? Create Update system Configure configuration and add gLite Create VO & CA Evaluate sites files repositories and set up VOMS TORQUE III a III ‘ III ‘ Configure worker nodes III ‘ + + Start services Prepare configuration Design Execution Descriptiveness Snapshotting Modularity Error handling Reusability Integration with lower-level tools Maintainability Human interaction Support for common patterns T. Buchert, L. Nussbaum Business workflows in distributed systems research 19 / 21

How BPM fulfills the requirements? Create Update system Configure configuration and add gLite Create VO & CA Evaluate sites files repositories and set up VOMS TORQUE III a III ‘ III ‘ Configure worker nodes III ‘ + + Start services Prepare configuration Design Execution Monitoring Descriptiveness Snapshotting Modularity Monitoring Error handling Reusability Instrumentation Integration with lower-level tools Maintainability Data analysis Human interaction Support for common patterns T. Buchert, L. Nussbaum Business workflows in distributed systems research 19 / 21

Summary To sum up, we: introduced problems of research in distributed systems, described our contributions: requirements for an experimentation engine, 1 a new approach to experimentation based on BPM, 2 showed how our approach meets the defined requirements. T. Buchert, L. Nussbaum Business workflows in distributed systems research 20 / 21

Future work In the near future, we plan to: implement a workflow-based experiment engine , evaluate it using a set of well-chosen experiments, find more analogies between BPM and experimental science. http://www.loria.fr/ ~buchert/ Thank you for your attention. T. Buchert, L. Nussbaum Business workflows in distributed systems research 21 / 21