snwagh/falcon-public

Weird communication results

khoaguin opened this issue · 2 comments

khoaguin commented

Hi, I am training LeNet on localhost for 15 iterations. I got the results like in the picture, where it shows the total communication is 900 Mb, but the communication for P0 is already 1673.28 MB. Is there anything wrong with this?

image

snwagh commented

I think there is some issue with the communication part (there is some issue with parallelization as well). Either it is wrapping around the bit size or something else is off. Can you please probe into this issue further?

khoaguin commented

Yes, I am looking into it. So for example, if I change the number of iterations to 1, the communications become normal again
image

  • 👍1