-
Notifications
You must be signed in to change notification settings - Fork 338
Add bisection test #203
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: master
Are you sure you want to change the base?
Add bisection test #203
Conversation
|
Can't you already do that running the sendrecv_perf test, and setting NCCL_TESTS_SPLIT_MASK=(nranks/2)-1? Sure, that only works with nranks being a power of two; maybe your code is more generic. |
|
Being able to run this test when nranks is not a power of two is actually important to us |
52ebfc6 to
42a079f
Compare
Add bisection to makefile Add bisection doc in performance.md
Add bisection to makefile Add bisection doc in performance.md
42a079f to
ff27d6e
Compare
|
The code is ready to merge from our side |
|
@x41lakazam Given the new NCCL_TESTS_SPLIT option allowing to split ranks in any way and no longer tied to powers of two, could you describe a use case where this allows to test performance in a new/better way? |
|
Oh, I wanted the cross rail change removed. Why did you restore it? |
|
@sjeaugey I didn't notice this change, does it permit bisection on any number of ranks ? |
|
Yes, if you set NCCL_TESTS_SPLIT="MOD X" with X=N/2 then it should be a bisection test. |
In a bisection, each rank is paired with one other rank called 'peer', they both send and receive N messages, this number is bound to the
agg_itersparameter.The selection of this 'peer' rank is defined by the
getPeerfunction:https://github.com/x41lakazam/nccl-tests/blob/bisection_test/src/bisection.cu#L19