Private bug reported:
CXL Push Writes (Ultra-Low Latency) is an advanced capability in the
Compute Express Link (CXL) ecosystem that enables devices (e.g.,
accelerators, SmartNICs, memory expanders) to directly push data into
host memory or cache-coherent domains with minimal latency.
Unlike traditional pull-based or host-mediated data movement, push
writes allow devices to initiate memory updates without requiring CPU
intervention or explicit read cycles. This significantly reduces latency
and improves efficiency for data-intensive and latency-sensitive
workloads.
This feature is particularly important in next-generation platforms
leveraging CXL 3.x fabrics, where large-scale disaggregated memory and
accelerator sharing require fast, coherent, and low-overhead data
exchange. Push writes enable near load/store-like semantics across the
fabric, improving responsiveness for real-time processing, AI/ML
pipelines, and high-performance networking.
In the Linux kernel, support for CXL Push Writes requires enhancements
across memory management, cache coherency handling, IOMMU, and device
driver models. While the underlying capability is hardware-driven, the
OS must correctly manage synchronization, ordering, and visibility of
device-initiated writes.
Feature Request:
Requested details to be enabled on OS:
Enable detection and enumeration of CXL push write capabilities.
Support device-initiated memory writes into host memory/coherent domains.
Ensure proper cache coherency and memory ordering for push-based updates.
Integrate with CXL.mem and CXL.cache subsystems in the OS.
Enhance IOMMU support for secure and efficient device-initiated writes.
Provide mechanisms for synchronization and visibility (e.g., barriers, notifications).
Support driver frameworks for push-based data movement models.
Expose configuration and status via sysfs/debugfs.
Enable performance monitoring and telemetry for push write operations.
Ensure compatibility with PCIe Gen6 and CXL 3.x fabrics.
Support virtualization environments (KVM/QEMU, SR-IOV, VFIO).
Provide validation and debugging tools for low-latency data paths.
Document usage models, constraints, and best practices.
Business Justification:
Enables ultra-low-latency data movement across CXL fabrics.
Reduces CPU overhead by eliminating unnecessary data polling and copies.
Improves performance for AI/ML, HPC, and real-time analytics workloads.
Enhances efficiency in disaggregated and composable infrastructure.
Supports next-generation accelerator and memory architectures.
Aligns with evolving CXL 3.x ecosystem and future data center designs.
References:
CXL 2.0 / 3.0 / 3.1 Specifications
Linux Kernel CXL Subsystem Documentation
PCIe Gen6 and Coherent Interconnect Documentation
Industry Whitepapers on Disaggregated Memory and Accelerator Fabrics
** Affects: linux (Ubuntu)
Importance: Undecided
Status: New
** Information type changed from Public to Private
--
You received this bug notification because you are subscribed to linux
in Ubuntu.
Matching subscriptions: Bgg, Bmail, Nb
https://bugs.launchpad.net/bugs/2146724
Title:
Request for CXL Support – Push Writes (Ultra-Low Latency) in Ubuntu
26.04
Status in linux package in Ubuntu:
New
Bug description:
CXL Push Writes (Ultra-Low Latency) is an advanced capability in the
Compute Express Link (CXL) ecosystem that enables devices (e.g.,
accelerators, SmartNICs, memory expanders) to directly push data into
host memory or cache-coherent domains with minimal latency.
Unlike traditional pull-based or host-mediated data movement, push
writes allow devices to initiate memory updates without requiring CPU
intervention or explicit read cycles. This significantly reduces
latency and improves efficiency for data-intensive and latency-
sensitive workloads.
This feature is particularly important in next-generation platforms
leveraging CXL 3.x fabrics, where large-scale disaggregated memory and
accelerator sharing require fast, coherent, and low-overhead data
exchange. Push writes enable near load/store-like semantics across the
fabric, improving responsiveness for real-time processing, AI/ML
pipelines, and high-performance networking.
In the Linux kernel, support for CXL Push Writes requires enhancements
across memory management, cache coherency handling, IOMMU, and device
driver models. While the underlying capability is hardware-driven, the
OS must correctly manage synchronization, ordering, and visibility of
device-initiated writes.
Feature Request:
Requested details to be enabled on OS:
Enable detection and enumeration of CXL push write capabilities.
Support device-initiated memory writes into host memory/coherent domains.
Ensure proper cache coherency and memory ordering for push-based updates.
Integrate with CXL.mem and CXL.cache subsystems in the OS.
Enhance IOMMU support for secure and efficient device-initiated writes.
Provide mechanisms for synchronization and visibility (e.g., barriers, notifications).
Support driver frameworks for push-based data movement models.
Expose configuration and status via sysfs/debugfs.
Enable performance monitoring and telemetry for push write operations.
Ensure compatibility with PCIe Gen6 and CXL 3.x fabrics.
Support virtualization environments (KVM/QEMU, SR-IOV, VFIO).
Provide validation and debugging tools for low-latency data paths.
Document usage models, constraints, and best practices.
Business Justification:
Enables ultra-low-latency data movement across CXL fabrics.
Reduces CPU overhead by eliminating unnecessary data polling and copies.
Improves performance for AI/ML, HPC, and real-time analytics workloads.
Enhances efficiency in disaggregated and composable infrastructure.
Supports next-generation accelerator and memory architectures.
Aligns with evolving CXL 3.x ecosystem and future data center designs.
References:
CXL 2.0 / 3.0 / 3.1 Specifications
Linux Kernel CXL Subsystem Documentation
PCIe Gen6 and Coherent Interconnect Documentation
Industry Whitepapers on Disaggregated Memory and Accelerator Fabrics
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2146724/+subscriptions
Комментариев нет:
Отправить комментарий