Abstract
An analysis of the performance of an Lanczos Eigenvalue Solver routine on a Cray T3E-900 MPP is described. The analysis based on the Amdahl law is presented for essentially non-sparse matrices. Efforts were done to reduce the sequential fraction of the code below 1%. It is proven that the collective communication of MPI can be very efficiently used on the Cray T3E to reduce the communication overhead of the parallel implementation. The results scale well up to 128 processing elements (PEs) for matrix dimensions of 8000, and is expected to scale better for higher dimensions.
References
M. Horoi and R. Enbody, Performance Analysis and Optimization of a Parallel Carbon Molecular Dynamic Code on a Cray T3E, submitted to ICPP98, URL: http://www.phy.cmich.edu/~horoi/post/icpp98.ps
G. M. Amdahl, Validity of Single Processor Approach to Achieving Large-Scale Computing Capability, Proc. AMPS Conf., pp 483-485, Reston, VA, 1967.
C. Lanczos, Vol. I, J. Res. Natl. Bur. Stand. 45, 255 (1950).
M. Horoi, B. A. Brown and V. Zelevinsky, Phys. Rev. C 50, R2274 (1994).
D. J. Kuck, High Performance Computing, Oxford University Press, 1996.
Author information
Authors and Affiliations
Editor information
Rights and permissions
Copyright information
© 1998 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
Horoi, M., Enbody, R. (1998). Efficient implementation of a Lanczos Eigenvalue solver on a Cray T3E-900. In: Sloot, P., Bubak, M., Hertzberger, B. (eds) High-Performance Computing and Networking. HPCN-Europe 1998. Lecture Notes in Computer Science, vol 1401. Springer, Berlin, Heidelberg. https://doi.org/10.1007/BFb0037229
Download citation
DOI: https://doi.org/10.1007/BFb0037229
Published:
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-540-64443-9
Online ISBN: 978-3-540-69783-1
eBook Packages: Springer Book Archive