The following is my input file for the BSE calculation. The GS calculations that inform the BSE seem to yield the same results («1% difference) and I can start with either a boron-10 or lithium STATE.OUT & EFERMI.OUT and get the same difference in exciton energies.

<input>

<title>BSE for Ge-H bulk</title>

<structure speciespath="/nfs/11/osu5259/EXCITING/exciting/species/">
<crystal scale="1.0">
<basevect>7.655137 0.0 0.0</basevect>
<basevect>-3.82757 6.629544 0.0</basevect>
<basevect>0.0 0.0 19.95477</basevect>
</crystal>
<species speciesfile="H3.xml" rmt="1.0000" >
<atom coord="0.3333344712623417 0.6666659952333239 0.7230181604860801" bfcmt="0.0 0.0 0.0" />
<atom coord="0.6666658947282684 0.3333355642013190 0.2230181376394524" bfcmt="0.0 0.0 0.0" />
<atom coord="-0.0000014047858665 -0.0000014989204292 0.8598346356745361" bfcmt="0.0 0.0 0.0" />
<atom coord="-0.0000010637550153 -0.0000023665415520 0.3598347357323680" bfcmt="0.0 0.0 0.0" />
</species>
<species speciesfile="Ge3.xml" rmt="1.9000" >
<atom coord="0.3333342808495061 0.6666677546212642 0.5752607854771313" bfcmt="0.0 0.0 0.0" />
<atom coord="0.6666676137995474 0.3333344202546681 0.0752607858377154" bfcmt="0.0 0.0 0.0" />
<atom coord="0.0000000779075186 0.0000000503072549 0.0071969085129899" bfcmt="0.0 0.0 0.0" />
<atom coord="0.0000000699936951 0.0000000808441511 0.5071959006397191" bfcmt="0.0 0.0 0.0" />
</species>
</structure>

<groundstate do="skip" ngridk="32 32 1" nempty="30" rgkmax="4" gmaxvr="20"/>

<xs xstype="BSE"
ngridq="4 4 1"
ngridk="4 4 1" vkloff="0.05 0.15 0.25"
nempty="30"
lmaxapwwf="3"
gqmax="2.0"
broad="0.0073499"
scissor="0.0101"
tevout="true" nosym="true">

<energywindow intv="0.0 1.0"
points="1200" />

<screening screentype="full"
nempty="215" />

<BSE bsetype="singlet"
nstlbsemat="1 30 1 30"
nstlbse="1 30 1 30"
aresbse="false"/>

<qpointset>
<qpoint>0.0 0.0 0.0</qpoint>
</qpointset>
</xs>

</input>

Thanks,
NikA

Any progress on the source of these differences?
Thanks, NikA

Dear NikA,

could you please post the input file that gives you problems, together with all the relevant information of the submission file (number of processors, number of nodes, etc.)?

Best regards,
Caterina Cocchi
(exciting team)

This is my input for the calculation that is crashing. The last files written in these calculations are the DIELTENS0_SCR.OUT and DIELTENS0_NOSYM_SCR.OUT (for exciting boron-10). The calculation crashes at the same point in the INFOXS.OUT file for both the lithium and boron-10 compilations, but I think the order in which the output files are written differs for these two versions.

<input>

<title>BSE for Ge-H bulk</title>

<structure speciespath="/nfs/11/osu5259/EXCITING/exciting/species/">
<crystal scale="1.0">
<basevect>7.522550 0.0 0.0</basevect>
<basevect>-3.761275 6.514719 0.0</basevect>
<basevect>0.0 0.0 41.8</basevect>
</crystal>
<species speciesfile="H.xml" rmt="1.0000" >
<atom coord="0.0000000000000000 0.0000000000000000 0.436741" bfcmt="0.0 0.0 0.0" />
<atom coord="0.3333333333333333 0.6666666666666667 0.612543" bfcmt="0.0 0.0 0.0" />
</species>
<species speciesfile="Ge.xml" rmt="1.9000" >
<atom coord="0.0000000000000000 0.0000000000000000 0.507203" bfcmt="0.0 0.0 0.0" />
<atom coord="0.3333333333333333 0.6666666666666667 0.542070" bfcmt="0.0 0.0 0.0" />
</species>
</structure>

<groundstate do="skip" ngridk="32 32 1" nempty="30" rgkmax="4" gmaxvr="20"/>

<xs xstype="BSE"
ngridq="4 4 1"
ngridk="4 4 1" vkloff="0.05 0.15 0.25"
nempty="30"
lmaxapwwf="3"
gqmax="2.0"
broad="0.0073499"
scissor="0.0101"
tevout="true"
nosym="true">

<energywindow intv="0.0 1.0"
points="1200" />

<screening screentype="full"
nempty="215" />

<BSE bsetype="singlet"
nstlbsemat="1 30 1 30"
nstlbse="1 30 1 30"
aresbse="false"
sciavbd="false"/>

<qpointset>
<qpoint>0.0 0.0 0.0</qpoint>
</qpointset>
</xs>

</input>

Thank you for your help,
NikA

Forgot submission details: the machine I work on have 20ppn, and given my q-mesh I have 10 q-points. I have tried 10 nodes with one process per node at various levels of threading (20,16,10) and 5 nodes with 2 processes per node (10 threads). The calculation freezes at the same place in both cases.
-NikA

Dear NikA,

is the job crashing for both exciting lithium and exciting boron-10? At which spot exactly is the calculation freezing in either case?

Best regards,
Caterina Cocchi
(exciting team)

The job crashes with boron-10 and produces NaN results for lithium.

For boron-10, the last task output in INFOXS.OUT is:

| EXCITING BORON-10 started for task 430 =
| MPI version using 10 processor(s) =
| Date (DD-MM-YYYY) : 19-03-2016 =

Info(xsinit): mapping screening-specific parameters

Info(findocclims): system has Kohn-Sham gap
Info(dfq): Gamma q - point: using momentum matrix elements for dielectric function
Info(dfq): number of G + q vectors (local field effects): 297
Info(dfq): lowest (partially) unoccupied state: 16
Info(dfq): highest (partially) occupied state : 15
Info(dfq): band-combination limits nst1, nst2, nst3, nst4: 15 216 216 15
Info(dfq): band-combination limits istl1, istu1, istl2, istu2: 1 15 16 231
Info(dfq): band-combination limits istl3, istu3, istl4, istu4: 16 231 1 15
Info(dfq): number of G + q vectors: 297
Info(df): Kohn Sham response function finished for q - point: 1

As previously noted, the output files stop writing after the DIELTENS0*.OUT are written.
For lithium, the tasks complete, but the EXCITON_SORTED*.OUT files are full of zeros and NaN values.
Thanks,
NikA

Dear NikA,

could you please check in your calculation whether you have partially occupied states? You can notice it by looking at the files EIGVAL_QMTXXX.

Another question: How much memory do your nodes have?

Best regards,
Caterina Cocchi
(exciting team)

No partially occupied states in the EIGVAL_*.OUT files. Nodes have 64 Gb memory but I can tell that unless the program is "protecting" the nodes by stopping before having an OOM problem, the issue is elsewhere: dynamic memory usage is monitored and is low when the code gets stuck.
Thanks,
NikA

Dear NikA,

I could successfully run your input file with exciting boron-10 using 10 processors distributed over 2 nodes.

As a general suggestion, when dealing with low-dimensional systems, it is advisable to converge the amount of vacuum in the non-periodic direction(s) starting from smaller amounts compared to the one shown in you input. On the other hand, a finer k-point mesh is definitely necessary to obtain reliable results.

Best regards,
Caterina Cocchi
(exciting team)

Thank you for your help! I was aware of the k-point convergence issue but wanted to successfully execute the code before getting too involved in using computer time. I will take your advice and investigate the vacuum convergence (this structure input was converged in another DFT code).
Thanks,
NikA