How to set num_samplers in DistDGL?

Questions
#1

I train GraphSAGE with ogbn-products dataset in two manchies.

it seems that the epoch times in the following four experiments do not differ much.

图片
图片2535×478 149 KB

how can I get a similar conclusion as Fig. 10 in paper DistDGL: Distributed graph neural network training for billion-scale graphs.

图片
图片1133×1059 61.6 KB

The time varies within a certain range.
Are 6.1 and 5.5 obtained by averaging？

#2

Setting num_samplers to 1 has some improvements, but it doesn’t seem to work as well for METIS

#3

launch.py --num_samplers x can set the number of samplers for each trainer. As for the detailed implementation of the paper you referred, please contact the author