Alex Granovsky
gran@classic.chem.msu.su
I've attached the complete input and output file for a very similar
job running using 16 physical cores of Intel(R) Xeon(R) E5-2630 v3-
based workstation. Please take a look at CPU/Wall clock times for
different stages.
Note, the MQCACI step is the only expensive stage that presently
runs using only a single core, and hence it can be rather slow.
This will be changed in future Firefly releases.
Conventional calculations with bigcas=1 would create huge disk files
and use several passes duplicating part of the work. Most likely they
would be more expensive than generator-direct ones. In addition, they require special I/O tuning options.
Hope this helps.
Best wishes,
Alex
On Wed Oct 18 '17 10:49pm, Dawid wrote
--------------------------------------
>Dear Alex,
>Thank you for your suggestion with $MCQGENS group. In fact, now I
>can do my check run without errors so hopefully I will manage to
>perform these calculations. Could you roughly estimate the computational
>cost increase due to using the same options in $MCQGENS as in
>your input compared to the default ones, please?
>Best wishes,
>Dawid
>On Wed Oct 18 '17 7:16pm, Alex Granovsky wrote
>----------------------------------------------
>>Dear Dawid,
>>very recently we have published a paper on fluorinated stilbenes
>>http://pubs.acs.org/doi/pdf/10.1021/jacs.7b09611
>>where we performed lots of XMCQDPT2/CAS(14,14) computations.
>>The most efficient and less memory-consuming way to perform such
>>computations with current Firefly is to use generators-direct code.
>>I'm attaching a part of a typical input file we used.
>>
>>
>>As to bigcas option and MQLMB2 memory demands, MQLMB2 allocates
>>memory using several sequential calls to memory management routines,
>>not a single call; This explains the behavior you are observing.
>>Hope this helps.
>>Kind regards,
>>Alex Granovsky
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>On Mon Oct 16 '17 1:03pm, Dawid wrote
>>-------------------------------------
>>>Dear Firefly Users,
>>>I want to perform XMCQDPT2 single-point calculations with
>>>CAS(14,14), i.e. very large CAS. I encounter numerous issues
>>>with memory. I have used
>>>$XMCQDPT2
>>>bigcas=1
>>>.
>>>.
>>>.
>>>$END
>>>to largely reduce memory requirements but it is still not enough.
>>>Now, I noticed that if I use either mwords=300 or mwords=320,
>>>I get following errors:
>>> ********** ROUTINE MQLMB2 ( 1) ***************************************
>>> ***** ERROR: MEMORY REQUEST EXCEEDS MAXIMUM.
>>> 301512972 WORDS REQUIRED 300000000 AVAILABLE.
>>> ADDRESS 0x093F7B87 HAS INITIATED PROGRAM ABORT BECAUSE OF FATAL ERROR(S)
>>>
>>>
>>>OR
>>>
>>>
>>> ********** ROUTINE MQLMB2 ( 1) ***************************************
>>> ***** ERROR: MEMORY REQUEST EXCEEDS MAXIMUM.
>>> 367923267 WORDS REQUIRED 320000000 AVAILABLE.
>>> ADDRESS 0x603D7B87 HAS INITIATED PROGRAM ABORT BECAUSE OF FATAL ERROR(S)
>>>
>>>
>>>So why do I have greater requirements when I give more?
>>>By the way, what can I do to decreases needed memory?
>>>I provide essential parts of input (I think so):
>>> $SMP
>>> np=4
>>> httnp=1
>>> smppar=.TRUE.
>>> $END
>>> $SYSTEM
>>> mwords=300
>>> timlim=36000
>>> $END
>>> $XMCQDPT
>>> bigcas=1
>>> inorb=1
>>> nstate=15
>>> wstate(1)=1,1,1,1
>>> edshft=0.02
>>> mxbase=30
>>> halloc=.TRUE.
>>> thrgen=1d-20
>>> genzro=1d-20
>>> threri=1.0d-20
>>> thrcon=1d-8
>>> $END
>>>Best wishes,
>>>Dawid
This message contains the 7319 kb attachment [ XMCQDPT2-CAS14,14.rar ] |