Fix the broadcast of norm2 in ED_GF_NORMAL #2
Assignees
Comments
beddalumia
added a commit
that referenced
this issue
Jan 10, 2023
This code has been increasingly diverging with respect to the single site one lately (or perhaps since ever). Not only on the key aspects allowing for treating clusters of sites, but in general in many other points of the shared infrastructure. I believe it may be time to pay a fair amount of this /technical debt/ and start working actively to reunite as many parts as we can, so to make it easier working on shared updates and eventually merge everything in a unique library (which I suppose is what we want to accomplish eventually... though not so easy to define a good way to do it for now). Here I start trying to bring in something that has been missing for a very long time, namely the "new" object oriented treatment of global (or pseudo global) sector quantities. Currently the CDMFT code features only a type for the sector *map* and not the type(sector) containing the map itself, alongside its spin-resolved dimensions, Nlanc, etc. This makes very uneasy to compare and import code from the single site code, which has built a lot on that baseline, in several years of development already. Hence let's start defining here this type(sector) too, excluding the fields that are obviously unused at the moment (phonon related, NONSU2&SUPERC...), and making build_sector and delete_sector work on these objects instead of the maps alone. This already triggers an inevitable cascade of changes for every call to those procedures. This directly allows to import the 'tridiag_Hv_sector' procedure, which has been needed for a long time to address issue #2, and now should enable a quick import/adaptation of the spin and charge impurity susceptibilities, from LIB_DMFT_ED. Should the refactory prove too heavy/problematic we would device a quicker fix to the issue and recode from scratch the susceptibilities, giving up the nice infrastructure, on a different branch. Hopefully this would not be the case.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
as it stands, the norm of vvinit is broadcast in this way, whenever we need to add/remove a particle
this creates an issue when a large number of cores is used, and the vector is small: only the cores in mpi_comm_group will get the norm, the others will get nothing. Then when they creare impG, the weights will all be 0. When they try to invert it in order to get Sigma, they will fail.
The solution can be
a) move the bcast after the delete_hv_sector line, where the communicator is reset
b) implement the new functions added in the single-site code (e.g. tridiag_Hv_sector_normal)
b) is preferred for uniformity and future-proof
The text was updated successfully, but these errors were encountered: