Scalable computing with parallel tasks

Abstract

Recent and future parallel clusters and supercomputers use SMPs and multi-core processors as basic nodes, providing a huge amount of parallel resources. These systems often have hierarchically structured interconnection networks combining computing resources at different levels, starting with the interconnect within multi-core processors up to the interconnection network combining nodes of the cluster or supercomputer. The challenge for the programmer is that these computing resources should be utilized efficiently by exploiting the available degree of parallelism of the application programs and by structuring the application in a way which is sensitive to the heterogeneous interconnect.In this article, we present an approach to structure the computations of an application as parallel tasks which can interact with other parallel tasks in communication phases. In particular, we consider how these parallel tasks can be mapped onto the computing resources provided by parallel clusters or supercomputers. We show that the scalability can be significantly improved by a suitable task-based organization and a corresponding structuring of the communication within tasks as well as between tasks. We evaluate the impact of different mappings of tasks to cores for different application programs on a variety of parallel machines.

Formalization of the concept of adaptive tasks mapping in the reconfigurable computers on FPGA

“…This allows to transform their graphs of algorithms to configure the most efficient task-oriented computing structures with a high processing speed in the reconfigurable environment. Computational algorithms with a mixed type of parallelism are presented by Macro Dataflow Graphs (MDGs) in the nodes of which macrofunctions (M-functions) are placed [22].…”

The method for providing quality of service time requirements in reconfigurable computing systems

“…The output computing application is a parallel program with mixed type of parallelism that is described by the programming model of M-tasks [10]. M-program is set by Macro Dataflow Graphs (MDG) [10] and macrograph tasks are placed in the vertices of it, and edges specify the relationship between vertices.…”

“…Mapping by levels is a common method of mapping tasks to computing structure. Herewith each level of the graph is sequentially mapped on computer system structure [10].…”

Design and implementation of a cloud computing service for finite element analysis

“…Calculating the number of different scheduling alternatives given a batch of independent tasks and computing resources is a known ''counting problem'' whose solution is given by the Stirling number of the second kind S(n, j), where n denotes the number of tasks and j the number of cores. The research model for hierarchical multiprocessor tasks (M-tasks) [29] also creates a data and control dependency graph among parallel tasks and schedules tasks to multi-core resources by layers (root-to-leaves) to assure overall progress. We did not focus on analysis of assembled (multi-part) systems and computational dependencies among them in this paper.…”