Container Service for Kubernetes:ack-cgpu component and changelog

cGPU module V1.5.20:

• Bug fix:

  Fixed a rare cGPU instance ID conflict that may occur during concurrent pod runs.

cGPU module V1.5.12:

• Bug fix:

  Fixed an issue where GPU memory isolation for some CUDA APIs failed with NVIDIA driver versions 535 and later.

cGPU module V1.5.9:

• Added support for ebmgn8ia instance types.

• Added support for cgroup v2 and Ubuntu 22.04.

cGPU module V1.5.7:

• Added support for gn8X instance types.

• Added support for NVIDIA driver versions 550 and later.

• Fixed a kernel panic issue that occurred on VMs after a node was destroyed.

cGPU module V1.0.9:

• Improved support for UVM devices.

• Adjusted the loading order of the wrapper and cGPU in the installation script to prevent cGPU from blocking the startup of ACK management pods.

• cGPU module V1.0.7: No updates.

• gpushare-device module: Added the gpushare=true label to enable sharing-only mode without memory isolation.

• cGPU module V1.0.7: No updates.

• gpushare-device module: Fixed an issue where ack-cgpu could not be installed on clusters running Kubernetes 1.24.

• cGPU module V1.0.7.

• gpushare-device module:

  • Fixed an issue where application pods failed to start because the GPU UUID was lost after an ECS instance reboot during shared GPU scheduling.

  • Added support for setting the cGPU policy via an environment variable.

cGPU module V1.0.7:

• Added support for additional regions, including Ulanqab, Guangzhou, and Chengdu.

cGPU module V1.0.6:

• Fixed an issue with residual memory after isolation, which previously caused memory limits to default to 8 GB after a failure.

• Resolved memory corruption issues that occurred after an out-of-memory (OOM) event in multi-process scenarios.

cGPU module V1.0.3:

• Added support for compilation and installation on kernel versions 5.1 and later and Alibaba Cloud Linux 3.2.

• Added support for the gn7 instance family. Adapted for CUDA 11 and NVIDIA driver versions 460 and later.

• Added policy3, which supports specifying computing power in percentages such as 50% or 30%. The minimum granularity is 5% with an error margin of ±5%.

• Fixed inaccurate memory display in multi-process scenarios.

• Fixed a GPU memory allocation issue with CUDA 11.4 and NVIDIA driver versions 470 and later.

cGPU module V0.8.13:

• Adapted for NVIDIA driver versions 450 and later.

• Expanded support to all available public cloud regions.

cGPU module V0.8.8:

• Initial release. Supported instance types: gn6i, gn6v, gn6e, gn5i, and gn5. Supported regions: China (Hangzhou), China (Shanghai), and China (Beijing).