Grid superscalara programming model for the grid

Resumen

During last years, the Grid has emerged as a new platform for distributed computing, The Grid technology allows joining different resources from different administrative domains and forming a virtual supercomputer with all of them. Many research groups have dedicated their efforts to develop a set of basic services to offer a Grid middleware: a layer that enables the use of the Grid. Anyway, using these services is not an easy task for many end users, even more if their expertise is not related to computer science. This has a negative influence in the adoption of the Grid technology by the scientific community. They see it as a powerful technology but very difficult to exploit. In order to ease the way the Grid must be used, there is a need for an extra layer which hides all the complexity of the Grid, and allows users to program or port their applications in an easy way. There has been many proposals of programming tools for the Grid. In this thesis we give an overview on some of them, and we can see that there exist both Grid-aware and Grid-unaware environments (programmed with or without specifying details of the Grid respectively). Besides, very few existing tools can exploit the implicit parallelism of the application and in the majority of them, the user must define the parallelism explicitly. Another important feature we consider is if they are based in widely used programming languages (as C++ or Java), so the adoption is easier for end users. In this thesis, our main objective has been to create a programming model for the Grid based on sequential programming and well-known imperative programming languages, able to exploit the implicit parallelism of applications and to speed them up by using the Grid resources concurrently. Moreover, because the Grid has a distributed, heterogeneous and dynamic nature and also because the number of resources that form a Grid can be very big, the probability that an error arises during an application's execution is big. Thus, another of our objectives has been to automatically deal with any type of errors which may arise during the execution of the application (application related or Grid related). GRID superscalar (GRIDSs), the main contribution of this thesis, is a programming model that achieves these mentioned objectives by providing a very small and simple interface and a runtime that is able to execute in parallel the code provided using the Grid. Our programming interface allows a user to program a Grid-unaware application with already known and popular imperative languages (such as C/C++, Java, Perl or Shell script) and in a sequential fashion, therefore giving an important step to assist end users in the adoption of the Grid technology. We have applied our knowledge from computer architecture and microprocessor design to the GRIDSs runtime. As it is done in a superscalar processor, the GRIDSs runtime system is able to perform a data dependence analysis between the tasks that form an application, and to apply renaming techniques in order to increase its parallelism. GRIDSs generates automatically from user's main code a graph describing the data dependencies in the application. We present real use cases of the programming model in the fields of computational chemistry and bioinformatics, which demonstrate that our objectives have been achieved. Finally, we have studied the application of several fault detection and treatment techniques: checkpointing, task retry and task replication. Our proposal is to provide an environment able to deal with all types of failures, transparently for the user whenever possible. The main advantage in implementing these mechanisms at the programming model level is that application-level knowledge can be exploited in order to dynamically create a fault tolerance strategy for each application, and avoiding to introduce overhead in error-free environments.