[ADOL-C] Large size advector slows down run time?
andrea.walther at uni-paderborn.de
Fri Mar 16 03:31:01 EDT 2018
Dear Choi Ji-Wung,
> Which constant should I change? OBUFSIZE, LBUFSIZE, VBUFSIZE, or TBUFSIZE?
to be on the safe side, I recommend to change all three.
> My program needs to handle dynamically changing indices of the data.
> I am wondering if there is any other way to do the task without using advector.
To judge this, we probably need more information about the application.
So we should discuss this without the mailing list in cc in an extra
email. This holds also truw for the topics mentioned below.
Kshitij, would it be possible to use a parameter vector for this case?
> 3) Is it possible to use advector in a sparse representation?
Kshitij, could you comment on that?
Thanks and best regards
> -----Original Message-----
> From: Andrea Walther [mailto:andrea.walther at uni-paderborn.de]
> Sent: Thursday, March 15, 2018 8:48 AM
> To: Choi Ji-Wung <ji-wung.choi at scania.com>; adol-c at list.coin-or.org
> Subject: Re: [ADOL-C] Large size advector slows down run time?
>> So it seems obvious to use advector.
>> But only constructing such size of advector slows down runtime,
>> advector vec(160000);
>> // vec is not even used anywhere.
>> // But, after constructing the variable, runtime of
>> IpoptApplication::OptimizeTNLP() that internally calls ADOL-C has become
>> significantly slower.
> yes that fits to our expectation since internally the book keeping for
> a new vector of size 160000 has to be done.
> You might check whether this leads to the storage of the internally
> generated traces of ADOL-C to disc. This would explain that you observe
> a severe slow down.
>> Do you have any suggestion to resolve this problem?
> If the traces are not written to disc then the only possibility is to
> somehow avoid the generation of that vector if it is not really necessary.
> If the traces are written to disc then an increase of the buffer sizes in
> will help. Then the traces are kept in main menory and therefore the
> access times are faster.
> Hope this helps. Otherwise just send an email again.
> Best regards
> Andrea Walther
Prof. Dr. Andrea Walther
Lehrstuhl fuer Mathematik und ihre Anwendungen
Institut fuer Mathematik
Warburger Str. 100
Email: andrea.walther at uni-paderborn.de
Phone: ++49 5251 602721
Fax: ++49 5251 603728
More information about the ADOL-C