A question about experiment results on paper about bns-gcn HOT 4 CLOSED

Hi @kcncoY, thank you for exploring BNS-GCN.

I'm confused by your first question because you are asking two completely different metrics :). Comm measures the communication time among workers, and memory stats reports the memory consumption of BNS-GCN. If you are asking how we measured the communication time, it is implemented in this file, and you can find an example about its usage here.

For correctly measuring the epoch time of BNS-GCN, please add --no-eval when you launch it. This ensures that no other tasks is launched between the end and start of the timer. The printed epoch time should be consistent among all workers.

Let me know if you have any other questions. Thanks!

from bns-gcn.

Thank you for your detailed answer！In the first question, I want to know the source of Epoch Comm (MB) rather than the communication time.

from bns-gcn.

As I remember, we did not directly measure communication volume (CV). Instead, we handly computed these results based on the number of transferred nodes and their embedding dimension. We also checked if the computed CV of BNS-GCN matches the CV measured by nvprof. You may check this document for understanding the usage of nvprof. Thanks.

from bns-gcn.

My problem has been resolved. Thank you very much for your response!

from bns-gcn.