Client Startup fails when n_workers more than physical cores
#1863
Comments
|
Interesting. Thank you for raising the issue. Can I ask you for more information about the system you're running this on? My current guess is a single box with 24 logical cores running Windows. Is this correct? Are you using a network file system at all? If you have the time I would be curious to see how things change with this pull request: #1852 If you are unfamiliar with git then you could test that by pip installing as follows: |
|
Also, does this happen every time or just sometimes? |
|
It happens every time. I restarted the computer as well just in case. The machine has 24 physical cores and 48 logical cores. I’m not using the file system (though in my trouble shooting I did try to use the To try out the pull request, is there a way to do so with conda? Or should I just use pip? |
|
You should just use pip. Pip is great at installing small source packages like this. I tend to use conda for big things (like numpy) and stable things (like normal dask releases) but pip for rapidly changing things (like git branches) |
|
And thanks for the extra information. I've seen this pop up a few times recently and I'm keen to track it down. If you're able to reliably reproduce it then I'm glad that you're around :) |
|
Also the machine is running Windows 10 64-bit. If you want/need the build information I can pass that along too.
… On Mar 27, 2018, at 16:34, Matthew Rocklin ***@***.***> wrote:
And thanks for the extra information. I've seen this pop up a few times recently and I'm keen to track it down. If you're able to reliably reproduce it then I'm glad that you're around :)
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub <#1863 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADh1D5LL_5CgrSS1oId8PAweS_4vkvErks5tiqJKgaJpZM4S9ina>.
|
|
I'm sorry, I gave you the wrong pip command above. I should have provided the following: Additionally I would also be curious about the behavior under this solution: |
|
Is there a preferred version I should use? |
|
Interesting. If you have some patience can I ask you to try each a few times, maybe in a for loop, and see if either of them breaks? from dask.distributed import Client
for i in range(100):
with Client() as client:
print(i) |
|
|
OK, that is very helpful. I appreciate your engagement here. Hopefully by working from one of those branches you'll be find short term. I'll merge one of those two for the next release, probably within a week or two. |
|
np, just to be clear, |
|
Short term I've gone with the other approach. There is a somewhat-rare failure on UNIX-based systems otherwise. Hopefully we'll be able to track it down soon. Thank you again for your collaboration here. Closing. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
If I try to start a local cluster with either
or
I get the first error message attached below.
If I start up a cluster with a number of workers equal to the number of physical cores (in my case 24) like this:
The I get the second error message below.
If I instead split the cluster and client creation across two ipython lines:
Then everything works fine.
Version numbers
First error message
Second Error Message
The text was updated successfully, but these errors were encountered: