This topic introduces cGPU and describes the benefits of cGPU Professional Edition by comparing it with cGPU Basic Edition.

Benefits of cGPU Professional Edition

Benefit Description
Supports graphics processing unit (GPU) sharing, scheduling, and memory isolation.
  • Supports GPU sharing, scheduling, and memory isolation on a one-pod-one-GPU basis. This is commonly used in model inference scenarios.
  • Supports GPU sharing, scheduling, and memory isolation on a one-pod-multi-GPU basis. This is commonly used to build the code to train distributed models.
Supports flexible GPU sharing and memory isolation policies.
  • Supports GPU allocation by using the binpack and spread algorithms.
    • Binpack: The system preferably shares one GPU with multiple pods. This applies to scenarios where high GPU utilization is required.
    • Spread: The system attempts to allocate one GPU to each pod. This applies to scenarios where the high availability of GPUs is required. The system attempts to avoid allocating the same GPU to different pod replicas of an application.
  • Supports GPU sharing without memory isolation. This applies to deep learning scenarios where applications are configured with user-defined isolation systems at the application layer.
  • Supports GPU sharing on multiple GPUs and memory isolation.
Supports comprehensive monitoring of GPU resources. Supports monitoring of both exclusive GPUs and shared GPUs.

Comparison between cGPU Basic Edition and cGPU Professional Edition

Feature cGPU Professional Edition cGPU Basic Edition
GPU sharing and scheduling on one GPU Supported Supported
GPU sharing and scheduling on multiple GPUs Supported Not supported
Memory isolation on one GPU Supported Supported
Memory isolation on multiple GPUs Supported Not supported
Monitoring and auto scaling of exclusive GPUs and shared GPUs Supported Supported
Node pools that support flexible policy configurations Supported. Allows you to create different GPU policies for a node pool. You can enable GPU sharing with or without memory isolation for a node pool. Supported. You can configure different GPU policies for a node pool. You can enable GPU sharing with or without memory isolation for a node pool. In addition, you can use the binpack or spread algorithm to allocate GPUs.
Allocate GPU memory to pods by using algorithms Supported. GPUs can be allocated by using the binpack and spread algorithms. You can choose binpack or spread to meet your business requirements. Supported. By default, GPUs are allocated by using the binpack algorithm.
Note Different cGPU editions are intended for different types of Kubernetes clusters:

GPU sharing solution by Alibaba Cloud

A key requirement of GPU sharing among multiple pods is to isolate the GPU memory and computing power that are allocated to each pod. When you run multiple containers on one GPU, the GPU resources are allocated to each container as required. However, if one container occupies excessive GPU resources, the performance of the other containers may be affected. To address this issue, many solutions have been developed in the computing industry. Technologies, such as NVIDIA virtual GPU (vGPU), NVIDIA Multi-Process Service (MPS), rCUDA, and vCUDA, all contribute to fine-grained GPU resource allocation.

The cGPU solution is provided by Alibaba Cloud. The cGPU solution uses the server kernel driver that is developed by Alibaba Cloud to provide more efficient use of the underlying drivers of NVIDIA GPUs. cGPU provides the following features:
  • High compatibility: cGPU is compatible with standard open source solutions, such as Kubernetes and NVIDIA Docker.
  • Ease-of-use: cGPU adopts a user-friendly design. To replace a Compute Unified Device Architecture (CUDA) library of an artificial intelligence (AI) application, you do not need to re-compile the application or create a new container image.
  • Stability: cGPU provides stable underlying operations on NVIDIA GPUs. API operations on CUDA libraries and some private API operations on cuDNN are difficult to call.
  • Resource isolation: cGPU ensures that the allocated GPU memory and computing capacity do not affect each other.

Based on cGPU, ACK enables GPU sharing and the scheduling of multiple tasks to one GPU. This enables GPU sharing and memory isolation for scheduled Kubernetes resources and the container runtime. This provides low-cost, reliable, and user-friendly GPU sharing and memory isolation for large scale business.