OK, issue solved. There is no performance problem, the current master was compiled with some shared libraries instead of static linking, while the old version had mainly static libraries compiled in. That caused the lag with the current version.
Sorry for making a fuss. Best, Marcin W dniu 05.08.2012 16:59, Barry Haddow pisze: > Hi Marcin > > I normally use valgrind (with the callgrind tool). > > cheers - Barry > > On 05/08/12 15:32, Marcin Junczys-Dowmunt wrote: >> Hi Barry, >> both version were run with one thread, but with multi-thread support >> compiled in. The slow-down is similar for 8 threads, just scaled >> proportionally, 170 seconds vs 135 seconds. >> >> I've noticed that effect while I was checking the performance of my >> phrase table after merging with master and first thought that it is my >> fault, but I have the same effect with the standard phrase table, as >> reported below. >> >> How exactly do you want me to profile the versions? >> Best, >> Marcin >> >> W dniu 05.08.2012 16:07, Barry Haddow pisze: >>> Hi Marcin >>> >>> Is this single-threaded or multi-threaded Moses? One possible issue is >>> that threads are now compiled in by default, so even if you run with a >>> single thread then you still pay some locking penalties. I'm not sure >>> that this should make the large difference that you observe though. You >>> can compile without threads to avoid the locking. >>> >>> I haven't noticed any significant slow-down, but we don't monitor >>> performance too carefully here (although we probably should). I'd be >>> interested to see a profile of the two versions that you tested, >>> >>> cheers - Barry >>> >>> On 05/08/12 11:27, Marcin Junczys-Dowmunt wrote: >>>> Hi, >>>> has anyone noticed that Moses seems to be about 20-30% slower than >>>> about >>>> a year ago? I happen to have a checkout from last September lying >>>> around. With the same settings, and the same data the translation of >>>> 1000 sentences takes around 925 seconds now, but 690 seconds with the >>>> September version. Translation results are the same. >>>> >>>> I use the same binary phrase table with around 200 million phrase >>>> pairs, >>>> no lexical reordering, the same 3-gram KenLm language model, but >>>> compiled to the corresponding format which changed between versions. >>>> The >>>> data is fully cached in the operational system read cache, so it seems >>>> this are processing issues. >>>> >>>> Has anyone else noticed something like that? >>>> Best, >>>> Marcin >>>> _______________________________________________ >>>> Moses-support mailing list >>>> [email protected] >>>> http://mailman.mit.edu/mailman/listinfo/moses-support >>>> >> > -- dr inż. Marcin Junczys-Dowmunt Uniwersytet im. Adama Mickiewicza Wydział Matematyki i Informatyki ul. Umultowska 87 61-614 Poznań _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
