Macro-dataflow Model Research Articles

Measuring the overhead of Intel C++ Concurrent Collections over Threading Building Blocks for Gauss–Jordan elimination

SUMMARYThe most efficient way to parallelize computation is to build and evaluate the task graph constrained only by the data dependencies between the tasks. Both Intel's C++ Concurrent Collections (CnC) and Threading Building Blocks (TBB) libraries allow such task‐based parallel programming. CnC also adapts the macro data flow model by providing only single‐assignment data objects in its global data space. Although CnC makes parallel programming easier, by specifying data flow dependencies only through single‐assignment data objects, its macro data flow model incurs overhead. Intel's C++ CnC library is implemented on top of its C++ TBB library. We can measure the overhead of CnC by comparing its performance with that of TBB. In this paper, we analyze all three types of data dependencies in the tiled in‐place Gauss–Jordan elimination algorithm for the first time. We implement the task‐based parallel tiled Gauss–Jordan algorithm in TBB using the data dependencies analyzed and compare its performance with that of the CnC implementation. We find that the overhead of CnC over TBB is only 12%– 15% of the TBB time, and CnC can deliver as much as 87%– 89% of the TBB performance for Gauss–Jordan elimination, using the optimal tile size. Copyright © 2012 John Wiley & Sons, Ltd.

Program Partitioning for NUMA Multiprocessor Computer Systems

Program partitioning and scheduling are essential steps in programming non-shared-memory computer systems. Partitioning is the separation of program operations into sequential tasks, and scheduling is the assignment of tasks to processors. To be effective, automatic methods require an accurate representation of the model of computation and the target architecture. Current partitioning methods assume today′s most prevalent models - macro dataflow and a homogeneous/two-level multicomputer system. Based on communication channels, neither model represents well the emerging class of NUMA multiprocessor computer systems consisting of hierarchical read/write memories. Consequently, the partitions generated by extant methods do not execute well on these systems. In this paper, we extend the conventional graph representation of the macro-dataflow model to enable mapping heuristics to consider the complex conununication options supported by NUMA architectures. We describe two such heuristics. Simulated execution times of program graphs show that our model and heuristics generate higher quality program mappings than current methods for NUMA architectures.

Mentat: An object-oriented macro data flow system

Mentat is an object-oriented macro data flow system designed to facilitate parallelism in distributed systems. Its underlying model of computation is the macro data flow model. The Macro data flow model is similar to the traditional, large grain data flow model with two differences: (1) Some macro actors are persistent actors that maintain state information between firings; (2) Program graphs are dynamic. Macro data flow programs are represented by dynamic structures called future lists that may be constructed and modified at run time. This paper discusses the motivation of the macro data flow model; the relationship between objects, actors, and tokens; the use of future lists; extensions to the programming language C++ for writing macro data flow programs; and the structure of a virtual macro data flow machine for executing macro data flow programs.