[SCIP] presolve time

[Marc Pfetsch]

Dear Chuen Teck,

presolving of SCIP-SDP is not running in parallel by default, so
changing OMP_NUM_THREAD should not have any effect. Consequently, most
of the fluctuations you observed have external reasons (e.g. the load of
your computer).

Please privately send me the input file (not to the list) and I will
have a look at the differences. Moreover, I would need to know your SCIP
version and how you measured the timing.

Best

Marc Pfetsch


On 02.04.19 03:50, Chuen Teck See wrote:
> Dear Scip,
> 
> 1) I encountered a discrepancy in presolve time. I used 2 methods of
> timing. The first is using verblevel = 5. The second is to put a timing
> around SCIP_CALL( SCIPpresolve(scip) ). The result is as follows.
> (Observe the portion in Red)
> 
> original problem has 79800 variables (0 bin, 79800 int, 0 impl, 0 cont)
> and 159601 constraints
>  159600 constraints of type <linear>
>       1 constraints of type <SDP>
> Reading Time: 0.71
> 
> solve problem
> =============
> 
> transformed problem has 79800 variables (0 bin, 79800 int, 0 impl, 0
> cont) and 159601 constraints
>  159600 constraints of type <linear>
>       1 constraints of type <SDP>
> 
> original problem has 239400 active (0.00187969%) nonzeros and 239400
> (0.00187969%) check nonzeros
> 
> presolving:
> (round 1, fast)       0 del vars, 159600 del conss, 0 add conss, 159600
> chg bounds, 0 chg sides, 0 chg coeffs, 0 upgd conss, 0 impls, 0 clqs
> clique table cleanup detected 0 bound changes
> 
> presolved problem has 79800 active (100%) nonzeros and 79800 (100%)
> check nonzeros
> 
> presolving (2 rounds: 2 fast, 1 medium, 1 exhaustive):
>  0 deleted vars, 159600 deleted constraints, 0 added constraints, 159600
> tightened bounds, 0 added holes, 0 changed sides, 0 changed coefficients
>  0 implications, 0 cliques
> presolved problem has 79800 variables (79800 bin, 0 int, 0 impl, 0 cont)
> and 1 constraints
>       1 constraints of type <SDP>
> transformed objective value is always integral (scale: 1)
> Presolving Time: 0.25
> SCIPpresolve(scip) time = 99.310370s
> 
> 2) Also, I used >export OMP_NUM_THREADS=numCore & obtain the following.
> (My cpu has 6 core)
> 6 core: SCIPpresolve(scip) time = 124s
> 4 core: SCIPpresolve(scip) time = 115s
> 2 core: SCIPpresolve(scip) time = 93s
> 1 core: SCIPpresolve(scip) time = 749s
> 
> 3) On a larger problem, 319600 variables (0 bin, 319600 int, 0 impl, 0
> cont) and 639201 constraints, the time is very long
> SCIPpresolve(scip) time on 2 core = 10538s
> 
> The scip I used is scipSdp (plugin developed by Tristan Gally). You can
> reproduce the behavior by compiling his plugin & I can provide the input
> file.
> 
> 
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
>