-
Notifications
You must be signed in to change notification settings - Fork 43
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
Scratch buffer copy-based implementation of ncclBcast / ncclBroadcast #417
Scratch buffer copy-based implementation of ncclBcast / ncclBroadcast #417
Conversation
@microsoft-github-policy-service agree company="Microsoft" |
Encountered a hang with rccl-test's ncclBcast runs. In rccl-test with ncclBcast, the buffer changes only for the root-gpu b/w test runs. So, the channel key changes only for the root gpu and remains the same for the other gpus. So, the root gpu starts to create a new sm channel while the other gpus do not create new sm channels. Hence, a hang occurs.
To fix this, we have to use copy-based implementation. This is the simplest copy-based implementation. Uses a static channel key with sm channels created using scratch buffer as the remote memories.