Heterogeneous HPC environments

Abstract

Directions of software technologies for innovative HPC environments are discussed according to the industrial user requirements for heterogeneous multidisciplinary applications, performance portability, rapid prototyping and software reuse, integration and interoperability of standard tools. The various issues are demonstrated with reference to the PQE2000 project and its programming environment SklE (Skeleton-based Integrated Environment). Modules developed by a variety of standard languages and tools are encapsulated into SklECL (SklE Coordination Language) structures to form the global application. A performance model associated to SklECL allows the static and dynamic tools to introduce a large amount of global optimizations without the direct intervention of the programmer. The paper discusses also some of the most critical issues in matching architecture and software technology, showing how the SklE environment, and its evolutions, can act as a solid framework in which innovative hardware-software systems for HPC can be studied and experimented.

COLTHPF, a run-time support for the high-level co-ordination of HPF tasks

“…This work will not begin from scratch. An optimizing compiler for SKIE (SKeleton Integrated Environment), a P 3 L-like co-ordination language [6], has already been implemented within the PQE2000 national project [24]. We only have to extend the compiler to support HPF tasks and provide suitable performance models which will allow the compiler to perform HPF program transformations and optimizations.…”

Constructing generic data-farm templates

“…A number of similar structured approaches to parallel programming exists; for a comparative review refer to [42]. Coincidentally closest to the PPF approach, in the sense that farms and pipelines occur amongst the skeletons, is the Pisa Parallel Programming Language, for example recently in [43]. The categorical data type (CDT) framework [42] for list programming is a related higher-level model which further step-backs from the control structures necessary for a particular parallel architecture and being polymorphic is without a preferred granularity.…”

Towards optimisation of openMP codes for synchronisation and data reuse

“…We translate OpenMP to tasks that may share data and map these tasks to a cc-NUMA platform. Most researchers who focus on task mapping problems have performed work to handle the mapping of processes on distributed systems [32] [30]. Our tasks may be finer grained than these.…”