graph-learn icon indicating copy to clipboard operation
graph-learn copied to clipboard
verified publisher icon alibaba

Do the PS and workers of tensorflow have to be one-to-one in Deploy Mode 2.

Open Zdm-Jdsc-lalala opened this issue 5 years ago • 3 comments

If I use Deploy Mode 2 , do the PS and workers of tensorflow have to be one-to-one? just like server and client. Thank you for your answer！ @baoleai @archwalker @alibaba-oss

Zdm-Jdsc-lalala avatar Oct 20 '20 10:10 Zdm-Jdsc-lalala

Make sure the number of workers is not less than that of PS. More than one workers may connect to ONE PS. And we suggest the number of workers can be divided by that of PS, to make load balance.

jackonan avatar Oct 21 '20 03:10 jackonan

Thank you for your answer! My main question is the use of Deploy Mode 2. If there are 10 machines in our school, under the condition of using Deploy Mode 2 : Can I use two machines as independent PS of TensorFlow, and the remaining 8 machines as one-to-one correspondence of server, client and TF worker? In other words, I have 2 separate TF PS, 8 servers, 8 clients and 8 TF workers. Thank you for your answer！ @jackonan

Zdm-Jdsc-lalala avatar Oct 21 '20 06:10 Zdm-Jdsc-lalala

I think this is what you need.

jackonan avatar Oct 28 '20 11:10 jackonan