v0.4.1

What's Changed

• Bump golang.org/x/net from 0.32.0 to 0.38.0 by @dependabot in #22

New Contributors

• @dependabot made their first contribution in #22

Full Changelog: v0.4.0...v0.4.1

v0.4.0

Feat

• Import Kubernetes client auth plugins
• support compatible with open-gpu-kernel-modules
• webhook support setting default runtimeClassName for vgpu pods

Fix

• use more accurate cgroup driver recognition
• multiple detail adjustments and bug fixes
• cuda api compatibility fix

Optimization

• container runtime compatibility optimization
• optimizing command line parameters help Information
• use hidden directories to mount host proc and host cgroup paths

Full Changelog: v0.3.1...v0.4.0

v0.3.1

Feat

• memory scaling allows reporting more device memory
• enhanced compatibility of dynamic library API
• metrics server add rate limiter

Fix

• fix dynamic library API error

Optimization

• optimize the core monitoring thread of the dynamic library

Full Changelog: v0.3.0...v0.3.1

v0.3.0

Feat

• webhook component supports setting default scheduling policies for pods
• scheduler extender component add SerialBindNode feature gate
• support optimal allocation mode based on device link topology
• webhook component supports setting default device topology mode

Fix

• fix configurable host manager dir path
• fix nvidia-smi topo numa info index out of range
• fix pod multi container scheduling device allocation exceeds actual limit #4
• fix inaccurate calculation of scheduler node score #5
• fix pod enable shareProcessNamespace container id extraction
• fix the issue with the interception library in CUDA version 12.8

Optimization

• optimize the numa topology algorithm to align numa as much as possible when allocating multiple GPUs
• optimize the multi node score performance of the scheduler
• optimize scheduler node resource collection
• enhance the interception effect of the library

Full Changelog: v0.2.1...v0.3.0

v0.2.1

Feat

• Provide helm charts deployment method
• Provide webhook admission service component

Fix

• Fix task interruption caused by device plugin initialization and installation #1

Full Changelog: v0.2...v0.2.1

v0.2

This version has undergone multiple improvements compared to v0.1

Feat

• Idle computing power of dynamic balancing equipment
• GPU device uses virtual memory after exceeding memory limit
• Rescheduling device allocation failed pod
• Include the memory used for graphic calculations in the memory limit

Fix

• Container parallel device allocation error #2
• Frequent calls to ListPodresource fail #3
• Fix warnings during compilation of CUDA interception library
• Fix some errors in the CUDA interception library

Optimization

• Optimize the efficiency of CUDA interception library
• Optimize the efficiency of multiple GPU card core limitations
• Adjust vGPU monitoring indicators
• Optimize GPU exclusive mode configuration
• Optimize the success rate of device allocation
• Move some feature switches to feature gate
• Refactoring to reduce some redundant code

Full Changelog: v0.1...v0.2

v0.1

Feat

• Support CUDA core and memory virtualization
• Support cgroupv1 and cgroupv2
• Support CUDA 12
• Support dual scheduling strategy
• Hard isolation of resources, the size of resources cannot be changed inside the container
• Support multiple VGPU resource monitoring indicators

Full Changelog: https://github.com/coldzerofear/vgpu-manager/commits/v0.1