Skip to content

Optimization of CPU hotspots with Progress/Control threads #7

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.

Sign up for GitHub

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

Jump to bottom
Closed
kpjeeja wants to merge 15 commits into from
Closed

Optimization of CPU hotspots with Progress/Control threads #7

kpjeeja wants to merge 15 commits into from

Conversation

@kpjeeja
Copy link

@kpjeeja kpjeeja commented Jan 9, 2026

Optimization of CPU hotspots with Progress/Control threads

    - This patch adds Adaptive sleep for progress Thread/control
    thread instead of while with timeout(0ms). if CQ completion does
    not thie progress Thread will keep spinning. Same applies for
    CM Thread.
    - Add Cq Burst read instead on 1 completion read.
    - in case of Eagain for send/write/read. add adaptive sleep
    delay to avoid udealy of some ms

Signed-off-by: Jeeja KP <[email protected]>

tsg- and others added 15 commits October 8, 2025 17:14
@tsg-
libfabric: add SYNAPSEAI support
9f4b19c
@tsg-
libfabric: Genericize discovery. Add verbs support.
14149d1
@tsg-
libfabric: better hw detection, FI_HMEM support
2a4391f 
- Enable smart auto-detection of device types
- Dual-path memory registration (GDR, FI_HMEM)
- Device type as custom param (priority: envvar, backend param,
  auto-detect)
@tsg-
libfabric: Add clear provider specific config
d674c07
@tsg-
libfabric: Add SynapseAI dmabuf support
b7c4b40
@tsg-
libfabric: bonding dev support in topo detection
d1f2b04
@tsg-
libfabric: assign bond dev to all GPUs v closest
74dca53
@tsg-
libfabric: deduplicate virtual device list
3917898 
Use the unique device list to resolve double free

Signed-off-by: Jerome Anand <[email protected]>
Signed-off-by: Tushar Gohad <[email protected]>
@tsg-
libfabric: Use desc-specific target offset
3e28f43 
This fixes a bug in multi-descriptor transfers where descriptors
point to different offsets within the same registered memory region.

Without this fix, RDMA reads always target offset 0. Should extract
each descriptor's specific target address instead.

Also impacted: Block-based transfers (Iteration N would read blocks
from iteration 0, etc), Partial buffer updates, etc.

Signed-off-by: Tushar Gohad <[email protected]>
@tsg-
libfabric: Unit tests for libfabric backend
22a4dfb 
- Create 16 descriptors at different 64KB offsets in a 1MB buffer
- Attempt RDMA transfers
@kpjeeja
Fix synapse dependency in meson
5b6c037 
Signed-off-by: Jeeja Kp <[email protected]>
@kpjeeja
Add support to nixl_test to verfiy libfabric backend
b4ad7ee 
- add libfabric backend to nixl test
- add libfabric Gaudi->Gaudi transfer test

Signed-off-by: Jeeja Kp <[email protected]>
@jeromean
Merge pull request intel-staging#1 from kpjeeja/jeeja_nixl_test
21e6f11 
Add nixl test for Gaudi with libfabric backend
@kpjeeja @jeromean
Add nixl install script
352045d 
ported from ofi branch,
commit f70b5d4
Author: Chendi Xue <[email protected]>

- changed the branch name from v0.6.0_OFI to libfabric

Signed-off-by: Jeeja KP <[email protected]>
@kpjeeja
optimize progressthread, cm thread and burst CQ and egain handling
b073533
@kpjeeja kpjeeja closed this Jan 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@kpjeeja @tsg- @jeromean