Programming models to support data science workflows
- Autores: Cristián Ramón Cortés Vilarrodona
- Directores de la Tesis: Rosa M Badia (dir. tes.), Jorge Ejarque Artigas (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
- Idioma: español
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Data Science workflows have become a must to progress in many scientific areas such as life, health, and earth sciences. In contrast to traditional HPC workflows, they are more heterogeneous; combining binary executions, MPI simulations, multi-threaded applications, custom analysis (possibly written in Java, Python, C/C++ or R), and real-time processing. Furthermore, in the past, field experts were capable of programming and running small simulations. However, nowadays, simulations requiring hundreds or thousands of cores are widely used and, to this point, efficiently programming them becomes a challenge even for computer sciences. Thus, programming languages and models make a considerable effort to ease the programmability while maintaining acceptable performance.
This thesis contributes to the adaptation of High-Performance frameworks to support the needs and challenges of Data Science workflows by extending COMPSs, a mature, general-purpose, task-based, distributed programming model. First, we enhance our prototype to orchestrate different frameworks inside a single programming model so that non-expert users can build complex workflows where some steps require highly optimised state of the art frameworks. This extension includes the @binary, @OmpSs, @MPI, @COMPSs, and @MultiNode annotations for both Java and Python workflows.
Second, we integrate container technologies to enable developers to easily port, distribute, and scale their applications to distributed computing platforms. This combination provides a straightforward methodology to parallelise applications from sequential codes along with efficient image management and application deployment that ease the packaging and distribution of applications. We distinguish between static, HPC, and dynamic container management and provide representative use cases for each scenario using Docker, Singularity, and Mesos.
Third, we design, implement and integrate AutoParallel, a Python module to automatically find an appropriate task-based parallelisation of affine loop nests and execute them in parallel in a distributed computing infrastructure. It is based on sequential programming and requires one single annotation (the @parallel Python decorator) so that anyone with intermediate-level programming skills can scale up an application to hundreds of cores.
Finally, we propose a way to extend task-based management systems to support continuous input and output data to enable the combination of task-based workflows and dataflows (Hybrid Workflows) using one single programming model. Hence, developers can build complex Data Science workflows with different approaches depending on the requirements without the effort of combining several frameworks at the same time. Also, to illustrate the capabilities of Hybrid Workflows, we have built a Distributed Stream Library that can be easily integrated with existing task-based frameworks to provide support for dataflows. The library provides a homogeneous, generic, and simple representation of object and file streams in both Java and Python; enabling complex workflows to handle any data type without dealing directly with the streaming back-end.
