gpudirect-tcpx: Update NCCL config manifest for GKE 1.34+ recommendat…#610
Merged
Merged
Conversation
jkru3
commented
May 28, 2026
| -x NCCL_GPUDIRECTTCPX_CTRL_DEV=eth0 \ | ||
| -x NCCL_GPUDIRECTTCPX_PROGRAM_FLOW_STEERING_WAIT_MICROS=500000 \ | ||
| -x NCCL_CROSS_NIC=0 \ | ||
| -x NCCL_ALGO=Ring,Tree \ |
Collaborator
Author
There was a problem hiding this comment.
For nccl-net 3.1.11+
jkru3
force-pushed
the
nccl-config-deprecation
branch
from
133cfa5 to
919680e
Compare
Jiaqicao257
approved these changes
Jun 1, 2026
…ions
This change updates the `nccl-config.yaml` ConfigMap manifest to remove deprecated environment variables and obsolete channel restrictions, aligning it with the official recommendations for the GKE 1.34+ TCPX stack.
Rationale for changes:
1. Removed `NCCL_GPUDIRECTTCPX_FORCE_ACK=0` & `NCCL_GPUDIRECTTCPX_TX_COMPLETION_NANOSLEEP=1000`
- Reason: These manual packet tuning variables are deprecated and completely ignored by the updated TCPX daemon (v2.0.15+) used in GKE 1.34. With the migration to COS 125 (Linux kernel 6.12+), the stack natively utilizes upstream Device Memory TCP (devmem TCP) for zero-copy transfers, making these custom daemon-level workarounds obsolete.
- Proof: These variables have been removed from the recommended configuration in the official Google Cloud GPUDirect-TCPX documentation:
https://cloud.google.com/kubernetes-engine/docs/how-to/gpu-bandwidth-gpudirect-tcpx#add-gpudirect-tcpx-manifests
2. Removed `NCCL_MAX_NCHANNELS=8` & `NCCL_MIN_NCHANNELS=8`
- Reason: Forcing the system to use exactly 8 channels is no longer recommended for H100 workloads running NCCL core 3.1.12+ (standard in GKE 1.34). Restricting the channel count prevents NCCL from dynamically selecting the optimal number of channels based on topology, which can artificially limit GPU network bandwidth.
- Proof: The official configuration guide no longer lists channel count limits, allowing NCCL to dynamically optimize itself:
https://cloud.google.com/kubernetes-engine/docs/how-to/gpu-bandwidth-gpudirect-tcpx#add-gpudirect-tcpx-manifests
These updates resolve the discrepancy where the manifest did not reflect the GKE 1.34 user guide recommendations.
…endations This change creates `nccl-config-latest.yaml` ConfigMap manifest to remove deprecated environment variables and obsolete channel restrictions, aligning it with the official recommendations for the GKE 1.34+ TCPX stack. Rationale for changes: 1. Removed `NCCL_GPUDIRECTTCPX_FORCE_ACK=0` & `NCCL_GPUDIRECTTCPX_TX_COMPLETION_NANOSLEEP=1000` - Rationale: These manual tuning parameters were workarounds for older, custom out-of-tree TCPX drivers. GKE 1.34 (COS 125) migrates to Linux Kernel 6.12+, which natively supports **Device Memory TCP (devmem TCP)**. The kernel's TCP stack now handles packet acknowledgment and zero-copy transfers natively, making these CPU-timing and socket-level workarounds obsolete. The new tcpx-daemon (v2.0.15) ignores these variables. - Proof (Linux Kernel v6.12 Merge): https://lore.kernel.org/netdev/20240831004313.3713467-1-almasrymina@google.com/ - Proof (Linux Kernel Documentation): https://www.kernel.org/doc/html/v6.12/networking/devmem.html 2. Removed `NCCL_MAX_NCHANNELS=8` & `NCCL_MIN_NCHANNELS=8` - Rationale: Setting these variables forces NCCL to bypass its internal, automatic topology-detection and channel-tuning algorithm. In newer NCCL versions (3.1.12+), this tuner is highly optimized to dynamically allocate the optimal number of channels (often up to 24 channels on A3/H100 nodes) to fully saturate the network bandwidth. Manually capping channels at 8 disables this optimization and acts as a performance bottleneck, which is recognized as a primary cause of communication regressions in distributed GPU training (and is actively asserted against in standard ML validation suites like Megatron-LM). - Proof (NVIDIA NCCL Tuning Documentation): Bypassing automatic channel selection is documented by NVIDIA as a manual override that should be avoided in production to allow topology-aware tuning: https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html These updates resolve the discrepancy where the manifest did not reflect the GKE 1.34 user guide recommendations.
…endations This change creates `nccl-config-latest.yaml` ConfigMap manifest to remove deprecated environment variables and obsolete channel restrictions, aligning it with the official recommendations for the GKE 1.34+ TCPX stack. Rationale for changes: 1. Removed `NCCL_GPUDIRECTTCPX_FORCE_ACK=0` & `NCCL_GPUDIRECTTCPX_TX_COMPLETION_NANOSLEEP=1000` - Rationale: These manual tuning parameters were workarounds for older, custom out-of-tree TCPX drivers. GKE 1.34 (COS 125) migrates to Linux Kernel 6.12+, which natively supports **Device Memory TCP (devmem TCP)**. The kernel's TCP stack now handles packet acknowledgment and zero-copy transfers natively, making these CPU-timing and socket-level workarounds obsolete. The new tcpx-daemon (v2.0.15) ignores these variables. - Proof (Linux Kernel v6.12 Merge): https://lore.kernel.org/netdev/20240831004313.3713467-1-almasrymina@google.com/ - Proof (Linux Kernel Documentation): https://www.kernel.org/doc/html/v6.12/networking/devmem.html 2. Removed `NCCL_MAX_NCHANNELS=8` & `NCCL_MIN_NCHANNELS=8` - Rationale: Setting these variables forces NCCL to bypass its internal, automatic topology-detection and channel-tuning algorithm. In newer NCCL versions (3.1.12+), this tuner is highly optimized to dynamically allocate the optimal number of channels (often up to 24 channels on A3/H100 nodes) to fully saturate the network bandwidth. Manually capping channels at 8 disables this optimization and acts as a performance bottleneck, which is recognized as a primary cause of communication regressions in distributed GPU training (and is actively asserted against in standard ML validation suites like Megatron-LM). - Proof (NVIDIA NCCL Tuning Documentation): Bypassing automatic channel selection is documented by NVIDIA as a manual override that should be avoided in production to allow topology-aware tuning: https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html Dry ran manifest with `kubectl apply --dry-run=client -f gpudirect-tcpx/nccl-config-latest.yaml` These updates resolve the discrepancy where the manifest did not reflect the GKE 1.34 user guide recommendations.
…files - Overwrites nccl-config.yaml with the verified GKE 1.34+ spec (removing deprecated vars like FORCE_ACK, MAX_NCHANNELS, etc., and enabling Ring,Tree). - Deletes the temporary nccl-config-latest.yaml which is now redundant. - Updates nccl-test-latest.yaml to point back to the standard nccl-configmap name.
jkru3
force-pushed
the
nccl-config-deprecation
branch
from
feaf51a to
0e5609b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Update NCCL config manifest for GKE 1.34+ recommendations