Breaking host-centric management of task-based parallel programming models
- Autores: Jaume Bosch Pons
- Directores de la Tesis: Carlos Álvarez Martínez (dir. tes.), Daniel Jiménez González (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2021
- Idioma: español
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Heterogeneous platforms had become popular to increase the computational power of the systems within a constrained power budget. They are present in several systems, from embedded platforms and mobile devices to high-end servers and clusters.
However, the co-processors are managed following a master-slave model where the general-purpose CPU drives the rest of elements. This management limits the system possibilities as not all application parts are suitable to be executed in an accelerator. This thesis presents different proposals to enhance the usage of co-processors in task-based parallel programming models, which are a powerful tool to easily program applications for heterogeneous platforms.
The first proposal enhances the task-based systems with an asynchronous, concurrent, and parameterizable behavior. The improvements go across the full-stack, from the programming model level down to the low-level communications used between the libraries and the co-processors. The evaluation shows that the implemented improvements boost the applications' performance as they can be easily tuned for the running platform.
The second proposal adds support for task spawn and synchronization in co-processors. The offloaded tasks can create child tasks that target other architectures or remain inside the co-processor. This allows the programmers to implement applications easily and effectively. The evaluation shows the efficiency of the proposal implementation in terms of latency and power consumption. The results show that applications can increase their performance and optimize their power consumption just moving the task spawn from the host threads to the co-processor. This is thanks to the low-latency task management inside the co-processors, which also reduces the communications between the host and the co-processor.
The third proposal extends task-based programming models with concepts of recurrent workloads. The regular task syntax has been extended with new clauses to label the recurrent tasks and provide the needed information to the runtime. The evaluation shows an application programmability increase thanks to the new syntax, which allows the specification of recurrent systems with much less code and better accuracy. Also, the direct management of task repetitions and periods in the co-processors allows an almost zero-latency management that is able to manage any task granularity.
