[Scip] Performance of Nehalem machine

[Anthony Presley]

Michael,

Thanks very much for the info.

Both systems are showing the same uname -ra, and are both using the
64-bit version of CentOS 5.4

I suppose it's possible that one has a slightly different version of the
kernel installed - though both are saying they have no updates to
install (with yum).

I actually copied the source from one machine to the other, and then
recompiled.

Very strange that investing a few thousand $ in a new machine, with
faster hardware, and more memory, results in slower performance.  It
makes it very, very difficult to benchmark.


--
Anthony

On Tue, 2010-04-20 at 20:20 +0200, michael.winkler at zib.de wrote:
> Hi Anthony,
> 
> unfortunately, it's always possible that some instances could run longer
> on even better processors.
> 
> It seems that the complete time difference appears to be the LP
> solvingtime, so maybe it's an CLP issue.
> 
> Another possibility is, that you have one machine with a 32 bit and
> another with a 64 bit system, which could also lead to such result. (Some
> compiler flags are different for 64 and 32 bit code in the SCIP
> makefiles.)
> 
> I hope that helps you a little, even if it's not your desired result.
> 
> Best regards, Michael
> 
> 
> > Hi all!
> >
> > I've tried solving an MPS problem (about 24MB in size), with 4011
> > variables and 65108 constraints.
> >
> > In trying to benchmark scip, I installed scip on to two servers, along
> > with CLP (from subversion).
> >
> > One machine has these characteristics: (machine1)
> > 	(2) Xeon 5570 Nehalem CPU's @ 2.93Ghz
> > 	24GB of RAM
> >
> > The other has these: (machine2)
> > 	(1) Xeon E5410 @ 2.33Ghz
> > 	8GB of RAM
> >
> > Neither machine is doing anything other than running scip (and whatever
> > background processing the kernel does).
> >
> > However, in this case, when getting to say, a gap of 4724.87%, machine1
> > looks like this:
> >
> >  time | node  | left  |LP iter| mem |mdpt |frac |vars |cons |ccons|cols
> > |rows |cuts |confs|strbr|  dualbound   | primalbound  |  gap
> > t 0.8s|     1 |     0 |     0 |  60M|   0 |   - |2866 |  46k|  46k|   0
> > |   0 |   0 |   0 |   0 |      --      | 1.681391e+07 |    Inf
> >  42.9s|     1 |     0 | 25711 |  82M|   0 | 353 |2866 |  46k|  46k|2866
> > |  46k|   0 |   0 |   0 | 3.481961e+05 | 1.681391e+07 |4728.86%
> > b43.3s|     1 |     0 | 26093 |  82M|   0 |   - |2866 |  46k|  46k|2866
> > |  46k|   0 |   0 |   0 | 3.481961e+05 | 1.680000e+07 |4724.87%
> >
> > And machine2 (to get to the same gap) looks like this:
> >
> >  time | node  | left  |LP iter| mem |mdpt |frac |vars |cons |ccons|cols
> > |rows |cuts |confs|strbr|  dualbound   | primalbound  |  gap
> > t 0.7s|     1 |     0 |     0 |  60M|   0 |   - |2866 |  46k|  46k|   0
> > |   0 |   0 |   0 |   0 |      --      | 1.681391e+07 |    Inf
> >  32.5s|     1 |     0 | 25711 |  82M|   0 | 353 |2866 |  46k|  46k|2866
> > |  46k|   0 |   0 |   0 | 3.481961e+05 | 1.681391e+07 |4728.86%
> > b32.8s|     1 |     0 | 26093 |  82M|   0 |   - |2866 |  46k|  46k|2866
> > |  46k|   0 |   0 |   0 | 3.481961e+05 | 1.680000e+07 |4724.87%
> >
> >
> > More specifically, the older, slower, processor solves (up to this
> > point) almost 10 seconds faster.
> >
> > I was expecting the opposite, out of a processor that is significantly
> > more expensive.
> >
> > Any ideas what's going on?
> >
> > Both scip's were built using "make LPS=clp" and that's it.  Coin was
> > built using "make".
> >
> > Both boxes are running an up to date CentOS 5.4
> >
> > Thanks!
> >
> >
> > --
> > Anthony
> >
> > _______________________________________________
> > Scip mailing list
> > Scip at zib.de
> > http://listserv.zib.de/mailman/listinfo/scip
> >
>