"I run it on the iphone for fun" -- Hieu, you're over the edge! However, I concur. Less than 500Mb per translation model is all you need for standard moses binary if you binarize everything. On a 4Gb Ram decoder machine with Raid0 hard disk arrays, we often run 4 or 5 translation models in parallel with passable performance.
I've also confirmed the discussion from a few months back about SSD's for swap/temp file space. Using 6-core Opterion with 16Gb Ram, 120 GB 2-disk SSD Raid0 for binarized storage gives me parity performance with running models in memory. Actually, tuning is slightly faster on SSD than Ram. I think because it's doesn't have the repeated delays of loading the models into memory each run. Tom On Fri, 26 Aug 2011 15:24:01 +0700, Hieu Hoang <[email protected]> wrote: > barry's right. > > Binarize the phrase table and the LM with irstlm or kenlm. Then just > look at the file sizes & add a few 100mb and that's your memory > requirement for adequate speed. > > You can run the phrase-based decoder in about 300mb if everything is > binarized. I run it on the iphone for fun :) > > (the chart decoder needs 1-2gb) > > On 26/08/2011 15:07, Barry Haddow wrote: >>> Ok, >>> i discovered that probably we can have a 64gb ram 8/12 cores >>> machine. >>> The requirements for translation are the same for the >>> training? >>> >>> I prepared two language models in binary format. And i >>> noticed that when the server is loading/translating it takes 89/90% >>> of >>> ram (actually the test environment has 4gb of RAM), and 10% of cpu. >>> But >>> when there aren't pending translation the memory used is 0%. >>> So for >>> translation machine i still need a 8/12 cores, or i can have a >>> "smaller" machine? >>> For translation what is important? Memory or CPU? >>> >>> And for example with 64gb ram, approximatively how many models can >>> i >>> load on the same machine (suppose we have models with >>> ~400'000/800'000 >>> sentences)? >>> >> >> Hi Ivan >> >> As far as ram is concerned, you need enough to load your model, any >> more won't >> make much difference, and any less then it will run impossibly slow >> due to >> swapping. >> >> If your data is processed in batches then you can benefit from >> having more >> CPUs and running multi-threaded decoding. >> >> I'm afraid I've no figures mapping training sentences to model size. >> I'd >> suggest that you run some experiments in your setup. >> >> cheers - Barry >> > _______________________________________________ > 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
