[Ipopt] More robust implementation of max_cpu_time option

Jonathan Hogg jonathan.hogg at stfc.ac.uk
Tue Apr 29 03:47:05 EDT 2014


We could be persuaded to add such a facility to HSL_MA97 if desired, 
though I think wall clock rather than cpu time might be more appropriate 
in this day and age. (We would still be at the mercy of lower level 
libraries such as BLAS/METIS however).

But before we dig through the linear solver code, it occurs to me that 
it might be appropriate to add a max_cpu_time test upon return from the 
linear solver. At present it's only check in the convergence test, and 
if my understanding of the code is correct, Ipopt could call the linear 
solver multiple times between convergence tests (eg if the inertia is 
wrong, or in the case of older codes if it runs out of memory).

This would however require some reengineering of the error handling that 
I don't know enough about to do safely?

Regards,

Jonathan.

On 28/04/14 23:06, Stefan Vigerske wrote:
> Hi,
>
> there is no such option.
> Ipopt has only very limit control over the solution of linear system, 
> as they are done by third-party software. These software packages 
> don't offer the possibility to pass on a timelimit.
>
> Stefan
>
> On 04/23/2014 05:54 PM, Roel De Coninck wrote:
>> Dear ipopt users,
>>
>> Is it possible to have a more strict implementation of the max_cpu_time
>> option?  The current implementation only checks the max_cpu_time 
>> during the
>> convergence step (
>> http://www.coin-or.org/Ipopt/documentation/node41.html#SECTION000112030000000000000 
>>
>> ).
>>
>> I have cases where I set the max_cpu_time, but it takes up to 10 times
>> longer before ipopt exits with the max_cpu_time exceeded status (eg 400s
>> instead of 25).  In these cases, it is always due to refactorizations
>> taking place.  These refactorizations are very slow, and it can take 
>> a long
>> time before the next convergence check.  Changing the refactorization
>> settings is not working, and would be a louzy workaround anyway.
>>
>> In short: is it possible for ipopt to exit (shortly) after the
>> max_cpu_time, whatever ipopt is doing at that moment?
>>
>> Thanks,
>> Roel
>>
>>
>>
>> _______________________________________________
>> Ipopt mailing list
>> Ipopt at list.coin-or.org
>> http://list.coin-or.org/mailman/listinfo/ipopt
>>
>
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt

-- 
Scanned by iCritical.


More information about the Ipopt mailing list