ELPH WF CACHE MB: Difference between revisions
Update related tags: add ELPH_WF_CACHE_PREFILL, ELPH_WF_COMM_OPT, ELPH_WF_REDISTRIBUTE; remove ELPH_MELS, ELPH_NQBATCH, ELPH_NSIM |
m Huebsch moved page Construction:ELPH WF CACHE MB to ELPH WF CACHE MB without leaving a redirect |
||
| (2 intermediate revisions by one other user not shown) | |||
| Line 1: | Line 1: | ||
{{DISPLAYTITLE | {{DISPLAYTITLE:ELPH_WF_CACHE_MB}} | ||
{{TAGDEF|ELPH_WF_CACHE_MB|real|1000}} | {{TAGDEF|ELPH_WF_CACHE_MB|real|1000}} | ||
| Line 8: | Line 8: | ||
Electron-phonon matrix elements are sandwiches of the form | Electron-phonon matrix elements are sandwiches of the form | ||
:<math>\langle \psi_{n\mathbf{k}} | \Delta V_{\mathbf{q}} | \psi_{m\mathbf{k}'} \rangle,</math> | :<math>\langle \psi_{n\mathbf{k}} | \Delta V_{\mathbf{q}} | \psi_{m\mathbf{k}'} \rangle,</math> | ||
where <math>\mathbf{k}' = \mathbf{k} + \mathbf{q}</math>. Because k-points are distributed across MPI ranks, the wavefunction <math>\psi_{n\mathbf{k}}</math> needed to form the bra may reside on a different rank than the one computing the matrix element. | where <math>\mathbf{k}' = \mathbf{k} + \mathbf{q}</math>. Because k-points are distributed across MPI ranks, the wavefunction <math>\psi_{n\mathbf{k}}</math> needed to form the bra may reside on a different rank than the one computing the matrix element. {{TAG|ELPH_WF_CACHE_MB}} sets the maximum memory (in megabytes) used to cache these remotely fetched wavefunctions locally, avoiding repeated inter-rank MPI communication. | ||
A separate cache for the PAW projections is also sized proportionally to | A separate cache for the PAW projections is also sized proportionally to {{TAG|ELPH_WF_CACHE_MB}}. | ||
The wavefunction cache works together with {{TAG|ELPH_WF_CACHE_PREFILL}} (default: <code>.TRUE.</code>), which pre-populates the cache before the main electron-phonon loop begins. When pre-fill succeeds, almost all subsequent wavefunction accesses are served from the cache without MPI communication. | The wavefunction cache works together with {{TAG|ELPH_WF_CACHE_PREFILL}} (default: <code>.TRUE.</code>), which pre-populates the cache before the main electron-phonon loop begins. When pre-fill succeeds, almost all subsequent wavefunction accesses are served from the cache without MPI communication. | ||
Latest revision as of 13:57, 18 June 2026
ELPH_WF_CACHE_MB = real
Default: ELPH_WF_CACHE_MB = 1000
Description: Maximum memory (in MB) allocated for caching wavefunctions during electron-phonon matrix element calculations.
| Mind: Available as of VASP 6.5.1 |
Electron-phonon matrix elements are sandwiches of the form
- [math]\displaystyle{ \langle \psi_{n\mathbf{k}} | \Delta V_{\mathbf{q}} | \psi_{m\mathbf{k}'} \rangle, }[/math]
where [math]\displaystyle{ \mathbf{k}' = \mathbf{k} + \mathbf{q} }[/math]. Because k-points are distributed across MPI ranks, the wavefunction [math]\displaystyle{ \psi_{n\mathbf{k}} }[/math] needed to form the bra may reside on a different rank than the one computing the matrix element. ELPH_WF_CACHE_MB sets the maximum memory (in megabytes) used to cache these remotely fetched wavefunctions locally, avoiding repeated inter-rank MPI communication.
A separate cache for the PAW projections is also sized proportionally to ELPH_WF_CACHE_MB.
The wavefunction cache works together with ELPH_WF_CACHE_PREFILL (default: .TRUE.), which pre-populates the cache before the main electron-phonon loop begins. When pre-fill succeeds, almost all subsequent wavefunction accesses are served from the cache without MPI communication.