That's strange, since as far as I could figure out, Cholesky factorization in cdiag.o seems to be called only in parallel diagonalizations - it's related to the decomposition of matrices over processors. You can trace the error from cdiag to a ScaLAPACK routine, can't recall its name, and take a look yourself. At the same time, when you parallelize over K-points, the matrices are diagonalized sequentially for each k-point by the corresponding CPUs.
Still, I remember having this problem with k-point parallelization, as well. It was solved (I think) by explicitly setting the BlockSize in the input file. Could be some sort of a bug, since BlockSize refers to parallelization over states, not k-points. Another flag that seem to help from time to time is Diag.Use2D .false., it calls a different ScaLAPACK routine. Of course, it would be very nice if the developers could look into this... 2007/6/2, Mariusz Krawiec <[EMAIL PROTECTED]>:
Hi there, Has anybody solved the problem of Cholesky factorisation in parallel siesta ? I got the same error on IA-64 when using Diag.ParallelOverK. When I run it on a single processor (even with mpirun command) everything works fine. All the best, Mariusz. ---------------------------------------------------------------------- Dr Mariusz Krawiec Institute of Physics Phone: +48 81 537 6146 M Curie-Sklodowska University Fax: +48 81 537 6191 pl. M. Curie-Sklodowskiej 1 E-mail: [EMAIL PROTECTED] 20-031 Lublin, Poland http://www.fnano.umcs.lublin.pl ---- ---- If you think of MS-DOS as mono, and Windows as stereo, then Linux is Dolby Pro-Logic Surround Sound with Bass Boost and all the music is free. - unknown ---- ----
