Hi, in the normal setup, the decoder reads the language model and translation model upon startup and then the only disk interaction is the input sentence and output sentence.
Since the models can get very large, there are ways to reduce the amount of RAM that is being used. Both of these have as drawback that the decoder runs slower. The language model can be partially kept on disk, if the IRST LM memory map used. The translation model can be kept completely on disk, if the binary translation table format is used. I hope this helps. -phi On Jan 18, 2008 5:04 PM, marco turchi <[EMAIL PROTECTED]> wrote: > Dear experts, > I'm running 4 instances of Moses on 4 different languages on the same > machine (4Gb, 2x 3.2GHz (2MB cache) P4 CPUs) and I have problems because > they take a lot of time, memory and CPU time... and I'd like to understand > where I can try to improve the machine... > > I have few points that i do not understand about how Moses works during the > decoding: > 1)RAM problem: Does Moses memory usage depend on the number of sentence that > I should translate? I can understand the difference between short and long > sentence, but I would not see the difference between translate 1k or 2k of > sentences in term of memory. I suppose that Moses gets 1 sentence, > translates it and releases the resource, and then passes to the next one. > Does Moses get into memory all the TM, LM and reordering table when it > starts or it accesses to the TM, LM and reordering table periodically on the > disk? > 2)I know for sure that the machine swaps memory to disk, but does Moses > access to the hard disk only for read the input sentence and write the > output sentence? or are there other r/w process during the decoding? > > > Thanks a lot > Marco > > > _______________________________________________ > Moses-support mailing list > [email protected] > http://mailman.mit.edu/mailman/listinfo/moses-support > > _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
