Delite runtime takes long time for scheduling when running on several threads

18 views
Skip to first unread message

Alexander Filippov

unread,
Mar 11, 2014, 9:58:18 AM3/11/14
to opt...@googlegroups.com
Good afternoon!

I have tried to run SVM example (found together with OptiML sources). After succesfully staging  and producing kernels and .deg file, I tried to run them on 8 scala threads. What I see is Delite runtime takes more then 1 hour to produce needed files in genedatedCache/runtime directory. As far as I understand, this directory should contain scheduled code which will be executed on hardware. So, I suspect that all this time Delite is busy with sheduling(i.e. splitting) the program into required number of threads; just because running on 1 thread goes very fast. 
Is this expected behaviour or I am doing something wrong? Is there any way to accelerate this process? Is this effect expected to disappear when running one more time due to reusing already generated (cached) code?

Thank you in advance,
Alexander Filippov

Kevin Brown

unread,
Mar 11, 2014, 11:31:36 PM3/11/14
to opt...@googlegroups.com
Hi, 
An hour is definitely significantly longer than I would expect; it takes about 30 seconds on my computer for SVM at 8 threads.  We fixed a bug not too long ago that was causing very long startup times for higher thread counts.  Are you up to date with the latest commits?  If not, you should try pulling the latest changes on the 'develop' branch and see if the situation improves.

--Kevin


--
You received this message because you are subscribed to the Google Groups "OptiML" group.
To unsubscribe from this group and stop receiving emails from it, send an email to optiml+un...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Alexander Filippov

unread,
Mar 21, 2014, 9:01:15 AM3/21/14
to opt...@googlegroups.com
Hi, Kevin,

Thank you for your help! Latest revision works fine!

best regards,
Alexander

среда, 12 марта 2014 г., 7:31:36 UTC+4 пользователь Kevin Brown написал:
Reply all
Reply to author
Forward
0 new messages