Hi Mike

Using a cache per-thread removes the need for locking on the cache, so 
should be faster. The negative effect on mosesserver was unexpected, but 
if you'd be able to make your threadpool changes available then that 
would be really helpful.

As to the general performance problems, our own tests 
(http://www.statmt.org/mosescore/uploads/Internal/D1.3_Moses_v2_Release_Notes.pdf)
 
showed v2.1.1 to be slightly faster than v1, however we did not test all 
possible configurations. If you could tell us a bit more about your 
setup then it would help us find out what the problem is.

cheers - Barry

On 26/06/14 14:26, Mike Ladwig wrote:
> The results I posted were from experiments with moses-cmd and not from 
> my thread pool version of mosesserver, although the results were similar.
>
> I looked the in the archives, and can find no explanation for the new 
> cache approach (global to per-thread).  What use case is improved with 
> this approach?
>
>
> On Thu, Jun 26, 2014 at 9:19 AM, Hieu Hoang <[email protected] 
> <mailto:[email protected]>> wrote:
>
>     Is the performance ok with the command line version? Can you push
>     your thread pool changes to a new branch.
>
>     I'll take a look at it when I get the chance
>
>
>     On 25 June 2014 16:25, Mike Ladwig <[email protected]
>     <mailto:[email protected]>> wrote:
>
>         I'm trying to move from the moses 1.x release to 2.x, but have
>         encountered large performance issues. On my workstation
>         (Scientific Linux 6.5), using the same spa-eng data to create
>         two systems I get performance roughly 3x slower on release 2.1.1.
>
>         I started by comparing moseserver between the 1.x and 2.x
>         releases.  After discovering the "single phrase cache per
>         thread" issue, I rewrote mosesserver using a thread pool but
>         only got a 10-20% improvement.
>
>         Thinking I might not really have fixed mosesserver, I tried
>         comparing unmodified moses-cmd speed between releases.   The
>         values are in words per minute for a 2000 line, 48k word file.
>
>                       1T         4T       8T
>         Rel 1:   4850   16492   19500
>         Rel 2:   1742    5324      6559
>
>         Any suggestions?
>
>         Regards,
>         mike.
>
>         _______________________________________________
>         Moses-support mailing list
>         [email protected] <mailto:[email protected]>
>         http://mailman.mit.edu/mailman/listinfo/moses-support
>
>
>
>
>     -- 
>     Hieu Hoang
>     Research Associate
>     University of Edinburgh
>     http://www.hoang.co.uk/hieu
>
>
>
>
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to