Repository navigation
fix(rocr): honor NUMA memory policy for host allocations - #14056
Merged
Merged
Conversation
…ions bind_mem_to_numa() unconditionally mbind()s host-pool allocations to the owning CPU agent's NUMA node, overriding an explicitly-set process memory policy (numactl -m / set_mempolicy). Skip the per-agent mbind when the process has set an explicit policy (get_mempolicy() != MPOL_DEFAULT) so the policy governs host placement. HSA_FORCE_MBIND=1 restores the legacy per-agent binding. No change when no policy is set. Relates to ROCM-32080; tracked by ROCM-33395.
✅ All Policy Checks Passed
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🚫 Please fix the failed policies before requesting reviews. The following policy checks failed:
The |
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
The new shortcut bypasses required placement checks and binding for NoSubstitute allocations, including APU device allocations.
1 open issue
What changed in this PR
This PR updates ROCr host-memory placement to honor explicit NUMA memory policies instead of always binding allocations to the owning CPU agent’s node.
Changes:
- Detect explicit memory policies and skip per-agent binding when one is set.
- Add
HSA_FORCE_MBINDto restore legacy binding behavior. - Preserve existing placement behavior when no explicit policy is set.
| File | Description |
|---|---|
projects/rocr-runtime/libhsakmt/src/openclose.c |
Read the HSA_FORCE_MBIND environment setting. |
projects/rocr-runtime/libhsakmt/src/libhsakmt.h |
Declare the shared binding override. |
projects/rocr-runtime/libhsakmt/src/globals.c |
Define and document the binding override. |
projects/rocr-runtime/libhsakmt/src/fmm.c |
Detect explicit NUMA policies and conditionally bypass binding. |
Review effort: Balanced
dayatsin-amd
approved these changes
Oct 26, 2026
Contributor
|
LGTM, nit pick. Please adjust your AI agent to use the "new" PR description format. You can use this PR as an example: |
The early return that defers host placement to an explicit process memory policy ran for every allocation, including strict (NoSubstitute) binds. The udmabuf APU path forces NoSubstitute=1 to pin host memory to the GPU's nearest CPU node for hardware locality; deferring there silently dropped that bind and returned success, violating the no-fallback contract. Gate the deferral on !NoSubstitute so only substitutable host-pool allocations (kernarg / signals / RPC) honor the policy, while strict binds keep their hard per-agent mbind and error handling. Rename the helper to reflect that Linux mempolicy is per-thread (numactl-set policy is inherited by every thread). Relates to ROCM-32080; tracked by ROCM-33395.
shwetagkhatri
approved these changes
Nov 4, 2026
shwetagkhatri
left a comment
shwetagkhatri
left a comment
Contributor
There was a problem hiding this comment.
Accepted and LGTM. Both points are addressed - the early return is now gated on NoSubstitute, and the helper is renamed to thread_has_explicit_mempolicy. Thanks Chris.
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
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.
Motivation
bind_mem_to_numa()unconditionally issues anmbind(MPOL_PREFERRED | MPOL_F_STATIC_NODES, <owning CPU agent's NUMA node>)on every host-pool allocation. An explicit-rangembindoutranks the process/task memory policy, sonumactl -m/set_mempolicy()are silently inert for ROCr host buffers: the runtime always wins, and HPC jobs that pin memory by policy cannot place GPU-accessed host memory (kernarg, completion signals, RPC buffers) where they asked. This is the explicit-policy half of ROCM-32080 (the no-policy default half — callers selecting CPU-agent-0 host pools — is fixed upstream in llvm-project PR #233099; ROCr's per-agent binding is correct there).Technical Details
Defer to an explicitly-set process memory policy: when the process has one, skip the per-agent
mbindand let the policy govern host placement. Behavior is unchanged when no policy is set.mbindwhenget_mempolicy()reports a non-default policy (masking off theMPOL_F_STATIC_NODES/MPOL_F_RELATIVE_NODESmode flags before the comparison). Returning0lets the allocation proceed under the process policy and is also correct for theNoSubstitutepath.HSA_FORCE_MBIND=1restores the legacy per-agent binding (escape hatch).HSA_ZFB/PM4_TARGET_XCC):hsakmt_force_mbindglobal inglobals.c, read once at init inopenclose.c:init_vars_from_env()viahsakmt_safe_env_to_int, declaredexterninlibhsakmt.h— no inlinegetenvin the alloc path.Note:
cpuset.memsis kernel-enforced andmbindcannot escape it, so no ROCr change is needed for the cgroup case.Issue Tracking
JIRA ID : ROCM-33395
Relates to ROCM-32080. Complements upstream llvm-project PR #233099 (the no-policy default case).
Test Plan
Validated on b06-2 (2-socket EPYC 9655 + 8× MI350X gfx950, ROCm 10.2), node-1 GPU with the process pinned to node 1, tracing allocations with
strace -f -e trace=mbind. Four cases covering the honor path, the legacy override, and no-regression:numactl -m 1numactl -m 1numactl -m 1+HSA_FORCE_MBIND=1Test Result
mbindHSA_FORCE_MBIND=1All four cases PASS.
Submission Checklist