This paper presents an implementation model for efficient state dependent processing on data flow machines, and some results from evaluation on NTT's data flow machine simulator. In this model, which is based on communicating processes with streams, each element of a stream is sent directly from an instance of a stream-producing function to a corresponding instance of a stream-consuming function. The order of elements in the stream is preserved by schemata for sequentializing instances of the functions. There is no need for data structure on memory for a stream. Therefore, it is expected that the same number of instances of a stream processing function as the number of elements in the stream are invoked and executed nearly in parallel. A nondeterministic processing is realized in the same framework. Evaluation results show that this model can attain about 80% improvement in speedup compared with a conventional model which exploits parallelism by a non-strict data structure constructor, eager evaluation of element data, and a memory with synchronization capability. It is also shown that a mutually exclusive processing in a sample program, i.e. a functional level simulator of a data flow machine as implemented by this model, does not constrict parallelism.
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Masaru TAKESUE, "A Model and Evaluation of State Dependent Processing on Data Flow Machines" in IEICE TRANSACTIONS on transactions,
vol. E71-E, no. 5, pp. 514-522, May 1988, doi: .
Abstract: This paper presents an implementation model for efficient state dependent processing on data flow machines, and some results from evaluation on NTT's data flow machine simulator. In this model, which is based on communicating processes with streams, each element of a stream is sent directly from an instance of a stream-producing function to a corresponding instance of a stream-consuming function. The order of elements in the stream is preserved by schemata for sequentializing instances of the functions. There is no need for data structure on memory for a stream. Therefore, it is expected that the same number of instances of a stream processing function as the number of elements in the stream are invoked and executed nearly in parallel. A nondeterministic processing is realized in the same framework. Evaluation results show that this model can attain about 80% improvement in speedup compared with a conventional model which exploits parallelism by a non-strict data structure constructor, eager evaluation of element data, and a memory with synchronization capability. It is also shown that a mutually exclusive processing in a sample program, i.e. a functional level simulator of a data flow machine as implemented by this model, does not constrict parallelism.
URL: https://global.ieice.org/en_transactions/transactions/10.1587/e71-e_5_514/_p
Copy
@ARTICLE{e71-e_5_514,
author={Masaru TAKESUE, },
journal={IEICE TRANSACTIONS on transactions},
title={A Model and Evaluation of State Dependent Processing on Data Flow Machines},
year={1988},
volume={E71-E},
number={5},
pages={514-522},
abstract={This paper presents an implementation model for efficient state dependent processing on data flow machines, and some results from evaluation on NTT's data flow machine simulator. In this model, which is based on communicating processes with streams, each element of a stream is sent directly from an instance of a stream-producing function to a corresponding instance of a stream-consuming function. The order of elements in the stream is preserved by schemata for sequentializing instances of the functions. There is no need for data structure on memory for a stream. Therefore, it is expected that the same number of instances of a stream processing function as the number of elements in the stream are invoked and executed nearly in parallel. A nondeterministic processing is realized in the same framework. Evaluation results show that this model can attain about 80% improvement in speedup compared with a conventional model which exploits parallelism by a non-strict data structure constructor, eager evaluation of element data, and a memory with synchronization capability. It is also shown that a mutually exclusive processing in a sample program, i.e. a functional level simulator of a data flow machine as implemented by this model, does not constrict parallelism.},
keywords={},
doi={},
ISSN={},
month={May},}
Copy
TY - JOUR
TI - A Model and Evaluation of State Dependent Processing on Data Flow Machines
T2 - IEICE TRANSACTIONS on transactions
SP - 514
EP - 522
AU - Masaru TAKESUE
PY - 1988
DO -
JO - IEICE TRANSACTIONS on transactions
SN -
VL - E71-E
IS - 5
JA - IEICE TRANSACTIONS on transactions
Y1 - May 1988
AB - This paper presents an implementation model for efficient state dependent processing on data flow machines, and some results from evaluation on NTT's data flow machine simulator. In this model, which is based on communicating processes with streams, each element of a stream is sent directly from an instance of a stream-producing function to a corresponding instance of a stream-consuming function. The order of elements in the stream is preserved by schemata for sequentializing instances of the functions. There is no need for data structure on memory for a stream. Therefore, it is expected that the same number of instances of a stream processing function as the number of elements in the stream are invoked and executed nearly in parallel. A nondeterministic processing is realized in the same framework. Evaluation results show that this model can attain about 80% improvement in speedup compared with a conventional model which exploits parallelism by a non-strict data structure constructor, eager evaluation of element data, and a memory with synchronization capability. It is also shown that a mutually exclusive processing in a sample program, i.e. a functional level simulator of a data flow machine as implemented by this model, does not constrict parallelism.
ER -