Dear QE users, To lessen I/O with neb.x, what should I do? The 'outdir' grows really fast. I want advice from experienced users or devs.
Issue: Although the image parallelism (neb.x -ni $NumOfNodes) is scaling well with the number of nodes even with gigabit ethernet, massive I/O to 'outdir' is observed. 1. Regardless of the existence of the wfcdir option, wavefunctions from each MPI processes (pwscf.wfc1, pwscf.wfc2, pwscf.wfc3, ... pwscf.wfc$ProcPerNode) are written under 'outdir' except for the first and the last image. Is this a bug? 2. When the 'disk_io' option is set to 'nowf', the neb calculation fails at the beginning of the second iteration because wavefunction files cannot be opened for writing. To execute properly, 'low' or above is necessary. I think this is expected by the devs, but warning at the beginning would be better. These issues can be reproduced on: a. QE 6.7, intel 19.5.281 (both MPI & compilers), CentOS 7, with LUSTRE a. QE 6.5, intel 19.5.281 (both MPI & compilers), CentOS 7, with LUSTRE b. QE 6.7, intel 19.5.281 (both MPI & compilers), CentOS 6, without LUSTRE c. QE 6.5, intel 18.1.038 (both MPI & compilers), CentOS 6, without LUSTRE ( a = Subsystem A in this page: https://www.hucc.hokudai.ac.jp/en/supercomputer/sc-overview/ ) According to the batch queue system on system A, it wrote 32.7GB/hour when 4 x 40 processes each. The command is "mpiexec.hydra -n 160 neb.x -ni 4 -i $INP >>$OUT". Again, my question is, what can I do to lessen I/O other than the 'disk_io' option and 'wfcdir' option? ---Sender--- Takahiro Chiba Undergraduate at Hokkaido University [email protected] -----
5itr_OHrot.neb.in
Description: Binary data
_______________________________________________ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list [email protected] https://lists.quantum-espresso.org/mailman/listinfo/users
