Dear Hieu, Thank you for you time and the clear answers.
>>> ok, if you are using plain text models, then this is the 1st difference to my results. I don't optimize the plain text models, I don't wanna wait for 13 minutes and this is not how most users use the decoder. Well, perhaps, but for me as a user it should not matter, loading the models happens just once, when the system is started. So in production this is not a big issue, only if one wants to start things up and shut them down periodically and quite often. >>> But if you see bad scaling in Moses2 with binary models, please let me know Is not that so that when the text models are loaded into memory they get binarized and then the binary model is just the memory snapshot of the loaded text model? I would expect so because otherwise one would need to use different data structures for storing binary and text models. Moreover, this would also mean different execution times on these two model types. So I suspect that having a text model should only influence the loading time of the model but not the decoding times... >>> If you do load the plain text files, you should check that it doesn't use up all memory and has to disk swap I have 5 times more memory than all the models take all together, so swapping is not an issue. >>> I know it's faster on 1 thread, it should be much faster on lots of threads. Yes, I was expecting the same results, but somehow I get a different trend .... so this is why I am asking. >>> no, just pre-load the binary files Ok, thanks! Kind regards, Dr. Ivan S. Zapreev > On Thu, Mar 16, 2017 at 3:11 PM, Hieu Hoang <[email protected]> wrote: > >> >> >> * Looking for MT/NLP opportunities * >> Hieu Hoang >> http://moses-smt.org/ >> >> >> On 16 March 2017 at 13:16, Ivan Zapreev <[email protected]> wrote: >> >>> Dear Hieu, >>> >>> Thank you for a prompt and detailed reply! >>> >>> >>> So your server has 20 cores (40 hyperthreads) and 16GB RAM? If >>> that's correct, then the RAM size would be a problem - you need as much RAM >>> as the total size of your models, plus more for working memory and the OS. >>> >>> The amount of memory is 256 Gb and not 16. There are a number of 16 Gb >>> plates installed. >>> To my knowledge the machine is not hyperthreaded but just has 40 cores, >>> although I am now getting a bit doubtful about that. >>> >> 256GB is good. 20/40 core/hyperthreads is not important for the moment, >> but you should find out exactly what it is >> >> >>> >>> >> Do you run Moses command line, or the server? My timings are based on >>> the command line, the server is a little slower. >>> >>> Both Moses and Moses2 are run in the console mode (not server). The >>> model loading time is excluded from the measurements. I could not manage to >>> get the asynchronous XML-RPC to work, so for my experiments that would be >>> as if I used Moses/Moses2 in a single-thread mode. Therefore I used the >>> command-line version. >>> >>> >>> Do you run Moses directly, or is another evaluation process running >>> it? Are you sure that evaluation process is working as it should? >>> >>> Moses is run from command time under the "time" command of Linux, and so >>> are other systems we used in comarison. We look at the runtime and not the >>> CPU times, but we perform a number of experiments to measure the average >>> times and control the standard deviations. >>> >>> >>> Do you minimise the effect of disk read by pre-loading the models >>> into filesystem cache? This is usually done by running this before running >>> the decoder cat [binary model files] > /dev/null >>> >>> Nop, we did not do pre-loading, for none of the tools but perhaps this >>> is not an issue as we just measure the average model loading times and >>> subtract them from the average run-time with decoding. So the model loading >>> times are excluded from the results. Our goal was to measure and compare >>> the decoding times and how they scale in the number of threads. >>> >> You're using the Probing phrase-table with integrated reordering model, >> and a binary KenLM, right? >> >> If so, the loading time will be minimal (1-2 secs) since the binary >> format just memory map the whole data but doesn't actually load them into >> memory. However, the overwhelming amount of the time taken for decoding >> will be page faults while doing LM and pt random lookups. >> >> It would be no surprise that decoding speed for Moses and Moses2 would be >> similar without pre-loading - they are looking up the same data. >> >> >>> >>> >>> it may take a while, but I can't replicate your results without it. >>> Alternatively, I can provide you with my models so you can try & replicate >>> my results. >>> >>> The experiments are run on an internal server which is not visible from >>> outside. I shall explore the possibilities of sharing the models, but I am >>> doubtful it is possible. The university network is very restricted. Yet, I >>> am definitely open to re-running your experiments. If possible. >>> >> i can make it available. But the results will be the same unless you sort >> out your pre-loading >> >>> >>> Kind regards, >>> >>> Ivan >>> >>> >> > > > -- > Best regards, > > Ivan > -- Best regards, Ivan <http://www.tainichok.ru/>
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
