Resumen de Explotació de paral.lelisme multinivell en sistemes multiprocessador i memòria compartida
This thesis is centered in the parallel programming for shared- memory architectures.
Current parallel shared-memory machines are becoming larger (from 32 up to 1024 processors), leading to very high programming complexity. Although several programming models for this architectures have been proposed, recently, OpenMP has been adopted as the reference standard for the majority of vendors.
One of the new features available in the OpenMP programming model is the definition of the nested parallelism. Typically the loop level parallelism is the most exploited source of parallelism in numerical applications.
Several recent studies have shown that with large systems, the loop level parallelism is not enough to feed the number of processors available in the architecture. Nested parallelism appears as the way to continue increasing performance.
This thesis studies the nested parallelism exploitation and proposes several mechanisms for having an efficient exploitation. The thesis contributions take form of two main extensions to the OpenMP programming model. First, the introduction of the thread groups concept in the model Second, a proposal for including in the language the possibility of coding explicit point-to-point thread synchronizations in the form of precedence relations among the work originated in the work-sharing constructs The thesis work includes a significant effort in the implementation of a powerful research platform in OpenMP topics. The platform is composed by a compiling system, supporting the majority of OpenMP constructs plus the extensions proposed in this thesis.
An OpenMP runtime has been designed, starting from nthe Nthlib thread runtime, supporting the Nanothread Programming Model. The Nthlib has been modified in order to be adapted to support the OpenMP programming model. The two proposals in the thesis have been implemented inside the NthLib thread library.
All the proposals have been evaluated with real a
