search

NVIDIA / nccl-tests

NCCL Tests
BSD 3-Clause "New" or "Revised" License
775 stars 226 forks source link

Why getBw don't have access to agg_iters ? #202

Closed x41lakazam closed 4 months ago

x41lakazam commented 4 months ago

Hey !

From my understanding getBw calculates the bandwidth using the number of data transfers. On the other side, agg_iters repeats the execution of the test and therefore the total number of transfers is multiplied by agg_iters.

However it doesn't seem to me that getBw receives information about agg_iters and the count argument is thus much smaller than the real number of elements transfered.

What am I missing ?

Thanks !

x41lakazam commented 4 months ago

My bad, it is included directly in the test time

https://github.com/NVIDIA/nccl-tests/blob/c6afef0b6f76ffc55d4172d971be6cf5a08a73a4/src/common.cu#L467-L470