> On 09 Aug 2016, at 1:00 pm, Thomas Dejanovic <thomas.dejano...@gmail.com> > wrote:
> I have a simulation sending packets through a system. It works very > well until we start to use some Xilinx primitives. If I break the > system down and leave out the module that has Xilinx primitives, the > simulation runs at least an order of magnitude faster and elapsed time > increases linearly with increasing simulation time. Once we > instantiate code that contains Xilinx primitives, doubling the > simulation time (i.e. doubling the number of packets sent) increases > the simulation time by about a factor of 6. > > i.e. - sending 128 packets takes about 2 minutes of sim time. 265 => > 12 minutes, 512 packets => ~72 minutes and 1024 packets takes longer > than I was willing to wait (jobs were killed after running for more > than 350 minutes). > > So my question is, has anyone else observed this behavior? An apparent logarithmic increase in simulation time seems related to memory leaks. There was an issue reported by Adrien Prost-Boucle around 9 March this year. https://mail.gna.org/public/ghdl-discuss/2016-03/index.html That was 'cured' by Tristan's commits c14f262, f4a612 and f1ddf16 around the 15th of March (SourceTree reports times on this side of the international dateline). https://mail.gna.org/public/ghdl-discuss/2016-03/msg00014.html > this issue is now fixed by my latest patches. > There was a threshold effect that prevented memory reuse and therefore > slowed down simulation. So the interesting questions become what revision (and potentially build date/commit) of ghdl, what platform and what Xilinx primitives/macros (and from what tool and version/release) are used when encountering the problem. Follow that by whether or not you can monitor memory utilization of your running ghdl simulation to determine whether or not there is a leak (should you be using a later built ghdl than those commits). Tristan has some Linux releases, llvm and mcode of several dates after the patches. https://github.com/tgingold/ghdl/releases _______________________________________________ Ghdl-discuss mailing list Ghdl-discuss@gna.org https://mail.gna.org/listinfo/ghdl-discuss
