Dynamic shared data in structured parallel programming frameworks

This work originates from the wish to simplify the coding of irregular applications within structured parallel programming environments. In these environments parallelism is exploited by composing “skeletons”, i.e. parallelism exploitation patterns. The skeletal approach has been proved to be effective, at least if application algorithms can be somehow expressed in terms of skeleton composition. However, in some cases our skeletal frameworks fail in providing the application programmer with convincing solutions both from ease of programming and performance viewpoints. Major lacks of expressivity have emerged in dynamic/irregular algorithms and applications that oddly access to large data sets. The first part of the thesis moves along this path, and reports all attempts we made to improve the effectiveness of environments’ compiler, static optimizer, and run-time support. The main goal of the thesis is to take a step further with respect to the achieved results. In particular we aim to defeat expressivity lacks emerged in skeletal languages approaching irregular problems and dealing with dynamic data structures. The basic idea consists in providing the application designer with a shared address space and a skeletal framework that enables and enforces the co-design of (shared) dynamic data structures and (parallel) algorithms. At this aim, a new skeletal programming environment based on shared address programming is proposed (i.e. eskimo). The language is as an extension of a “host” language (i.e. the C language). eskimo is conceived to be a framework to experiment how to support dynamic data structures in a skeletal framework. Its run-time support is based on a software distributed shared memory, and allows the programmer to freely access data items in the shared memory. eskimo is designed to match the hooks offered by ASSIST, thus to be experimented within. Notably eskimo is not yet another DSM, rather it relies on DSM already known technologies to experiment the co-design of dynamic data structures and parallel programming patterns enforcing locality in the distributed memory access. eskimo has been designed and developed from scratch. eskimo run-time support exploits multithreading, dynamic data-driven scheduling, and is very tolerant with respect to standard POSIX programming framework. It will be released as open source package.