Abstract
Parallel systems for OPS5 have been developed. In such systems, parallel implementations of the Rete algorithm are adopted because the number of pattern matchings is minimized in the original algorithm. However, conventional approaches for parallel Rete algorithms require special hardware to cope with the dynamic process allocation and frequent communication. Dedicated machines are necessary for such methods.
In this paper, a new parallel execution method for general purpose multiprocessors is proposed. In this method, nodes in the Rete network are statically mapped onto a small number of processors (PUs) before execution. Since processes are not migrated and are not allocated dynamically, data structures can be designed so as to minimize the communication between PUs. Also the synchronization and process control can be simplified.
The key for the method is static mapping algorithms. Since the characteristics of the Rete network are completely different from the data dependency graph which is suitable for extracting inherent parallelism in general programming languages like FORTRAN, the conventional static mapping techniques cannot be applied. Some new mapping strategies are discussed and evaluated.
Although the method is utilized for most types of multiprocessors, an optimized architecture which is called MANJI-II is designed and proposed. MANJI-II supports quick transfer of tokens in the Rete network, by using a simple multicast mechanism, and non-cached shared memory. Although MANJI-II provides almost the same performance as that of dedicated architecture including the previous version of MANJI, the total cost of MANJI-II is extremely low.
Preview
Unable to display preview. Download preview PDF.
References
C.L.Forgy: OPS5 User's Manual, Dept. of Comp. Sci., Carnegie-Mellon Univ. (1981).
C.L. Forgy: Rete: A fast algorithm for the many pattern/many object pattern match problem, Artificial Intelligence, vol. 19, pp. 17–37 (1982).
A.Gupta, C.L.Forgy, A.Newell and R.Wedig: Parallel algorithms and architectures for rule-based systems, Proc. of 13th ISCA, pp. 28–37 (1986).
A. Gupta, C. L. Forgy, D. Kalp, A. Newell, and M. Tambe: Parallel OPS5 on the Encore Multimax, ICPP, 1988, pp. 271–280
A. Gupta, M. Tambe:Suitability of Message Passing Computers for Implementing Production Systems, AAAI 88, 1988, pp. 687–692
M.A.Kelly, R.E.Seviora: A Multiprocessor Architecture for Production System Matching, AAAI, pp. 36–41 (1987).
J. Miyazaki, H. Amano and H. Aiso: MANJI:An architecture for production systems, Proc. of 20th Hawaiian International Conference on System Sciences, vol. 1, pp. 236–244 (1987).
J.Miyazaki, H.Amano, K.Takeda and H.Aiso: A shared memory architecture for MANJI production system machine, IWDM'87, pp. 517–531 (1987).
H. Kasahara, S. Narita: Practical multiprocessor scheduling algorithms for efficient parallel processing, IEEE trans. on Comp., vol. c33 No11, pp. 1023–1029 (1984).
H.Amano: RSM:A communication mechanism for multiprocessors, Proc. 2nd of International Conference on Computers and Applications, pp. 149–156 (1987).
H.Amano, T.Boku, T.Kudoh and H.Aiso: (SM)2-II: A new version of the Sparse Matrix Solving Machine, Proc. of 12th ISCA, pp. 100–107 (1985).
H.Amano, T.Terasawa and T.Kudo: Cache with synchronization mechanism, To appear in IFIP CONGRESS'89
Author information
Authors and Affiliations
Editor information
Rights and permissions
Copyright information
© 1989 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
Miyazaki, J., Takeda, K., Amano, H., Aiso, H. (1989). A new version of a parallel production system machine, MANJI-II. In: Boral, H., Faudemay, P. (eds) Database Machines. IWDM 1989. Lecture Notes in Computer Science, vol 368. Springer, Berlin, Heidelberg. https://doi.org/10.1007/3-540-51324-8_44
Download citation
DOI: https://doi.org/10.1007/3-540-51324-8_44
Published:
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-540-51324-7
Online ISBN: 978-3-540-46197-5
eBook Packages: Springer Book Archive