I keep this bash script to run stuff on the CPU. But, for spacy in my case, 4 threads have always proved to be the fastest.pic.twitter.com/zw2YhPw6uT
This thread is very relevant for anyone running parallel experiments on a CPU with Pytorch - setting the OMP_NUM_THREADS=1 and MKL_NUM_THREADS=1 gives me a 27x speedup when running 8 parallel experiments on 8 cores.https://twitter.com/honnibal/status/1067920534585917440 …
If you try out the new spacy-nightly (v2.1.0a3), you might be surprised to see it's single-threaded. This actually took a tonne of work! I've spent many hours getting the Blis linear algebra routines into a stand-alone, wheel-installable package. Why? A thread on threads 
1/10
1/10
Show this thread
-
-
-
Right - my comment only applies when you are also have top level experiment parallism- for single experiments, using more threads is faster. It's just that when you have multiple experiments, they stomp on each other if they share resources
-
That’s right. I try to make sure that my top level parallelism * num-threads-per-job doesn’t exceed the available resources. And in that case I seem to get the most out of spacy with 4 threads. But that might just pertain to my use case.
End of conversation
New conversation -
-
-
I've seen a similar thing with tensorflow on cpu, 5x speedup for certain models with interop threads set to 1...amongst other things http://blog.tabanpour.info/projects/2018/09/07/tf-docker-kube.html …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.