High-level compiler analysis for openmp
- Autores: Sara Royuela Alcázar
- Directores de la Tesis: Xavier Martorell Bofill (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2018
- Idioma: español
- Tribunal Calificador de la Tesis: Edmond Schonberg (presid.), Jesús José Labarta Mancho (secret.), Michael Klemm (voc.)
- Programa de doctorado: Programa de Doctorado en Arquitectura de Computadores por la Universidad Politécnica de Catalunya
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Nowadays, applications from dissimilar domains, such as high-performance computing and high-integrity systems, require levels of performance that can only be achieved by means of sophisticated heterogeneous architectures. However, the complex nature of such architectures hinders the production of efficient code at acceptable levels of time and cost. Moreover, the need for exploiting parallelism adds complications of its own (e.g., deadlocks, race conditions,...). In this context, compiler analysis is fundamental for optimizing parallel programs. There is however a trade-off between complexity and profit: low complexity analyses (e.g., reaching definitions) provide information that may be insufficient for many relevant transformations, and complex analyses based on mathematical representations (e.g., polyhedral model) give accurate results at a high computational cost.
A range of parallel programming models providing different levels of programmability, performance and portability enable the exploitation of current architectures. However, OpenMP has proved many advantages over its competitors: 1) it delivers levels of performance comparable to highly tunable models such as CUDA and MPI, and better robustness than low level libraries such as Pthreads; 2) the extensions included in the latest specification meet the characteristics of current heterogeneous architectures (i.e., the coupling of a host processor to one or more accelerators, and the capability of expressing fine-grained, both structured and unstructured, and highly-dynamic task parallelism); 3) OpenMP is widely implemented by several chip (e.g., Kalray MPPA, Intel) and compiler (e.g., GNU, Intel) vendors; and 4) although currently the model lacks resiliency and reliability mechanisms, many works, including this thesis, pursue their introduction in the specification.
This thesis addresses the study of compiler analysis techniques for OpenMP with two main purposes: 1) enhance the programmability and reliability of OpenMP, and 2) prove OpenMP as a suitable model to exploit parallelism in safety-critical domains. Particularly, the thesis focuses on the tasking model because it offers the flexibility to tackle the parallelization of algorithms with load imbalance, recursiveness and uncountable loop based kernels. Additionally, current works have proved the time-predictability of this model, shortening the distance towards its introduction in safety-critical domains.
To enable the analysis of applications using the OpenMP tasking model, the first contribution of this thesis is the extension of a set of classic compiler techniques with support for OpenMP.
As a basis for including reliability mechanisms, the second contribution consists of the development of a series of algorithms to statically detect situations involving OpenMP tasks, which may lead to a loss of performance, non-deterministic results or run-time failures.
A well-known problem of parallel processing related to compilers is the static scheduling of a program represented by a directed graph. Although the literature is extensive in static scheduling techniques, the work related to the generation of the task graph at compile-time is very scant. Compilers are limited by the knowledge they can extract, which depends on the application and the programming model. The third contribution of this thesis is the generation of a predicated task dependency graph for OpenMP that can be interpreted by the runtime in such a way that the cost of solving dependences is reduced to the minimum.
With the previous contributions as a basis for determining the functional safety of OpenMP, the final contribution of this thesis is the adaptation of OpenMP to the safety-critical domain considering two directions: 1) indicating how OpenMP can be safely used in such a domain, and 2) integrating OpenMP into Ada, a language widely used in the safety-critical domain.
