On Wed, 2007-11-28 at 00:31 +0300, Andrew Makhorin wrote: > > We have a (what I consider) to be fairly simple model, which performs > > very well for small - medium data sets (between 500 and 2500 lines of > > data). However, we recently tried to use the mod file for a larger data > > set (about 3200 lines of data). Note: We're not using glpk to solve our > > Integer Program, we use it to create an MPS file, which is then solved > > by another solver, so we use the --check option. Using this method, we > > have an MPS file generated in under 20 seconds. > > > Using a faster Core 2 Duo, 2GB of RAM, and 2GB of swap, glpk ran fine > > until our last rule. In the last rule, memory sky rocketed from 400MB > > for the glpk process to more than 4GB, resulting in a system load of > > over 20, and glpk eventually quit with a malloc() error. It took over > > 60 minutes to get to this point. > > > So, I moved it to our production system, which has 4GB of RAM and 4GB of > > swap, and the same thing occurred. It took over 40 minutes to get to > > this point. > > > Our original data sets use about 200 - 300MB, in total. > > > Any ideas on how we might go about speeding up the creation of the MPS > > file .... and bringing down the memory footprint? > > Most probably this happens due to huge amount of data, because > internal model representation used in the MathProg translator has > relatively high overhead expenses. > > Is there an excessive swapping on the generation phase? > > Which is the model size (estimated) for successful and unsuccessful > runs, resp.? >
_______________________________________________ Help-glpk mailing list [email protected] http://lists.gnu.org/mailman/listinfo/help-glpk
