基于Docker的高性能并行异构系统设计1

基于 Docker 的高性能并行异构系统设计与实现

摘要

在高性能计算系统中构建和部署基于图形处理加速器的软件环境是一项极

具挑战性的任务。高性能应用程序必须可靠地运行在多个平台和环境中，并在解

决复杂的软件堆栈依赖关系的同时充分利用图形处理加速器的资源。容器通常用

于在多台机器上无缝部署基于 CPU 的应用程序。有了这个用例，容器就是硬件无

备便携式和跨平台的特点，我们使用的解决方案是分离图像与英伟达驱动程序之

间的联系。启动目标机器上的容器时，将安装所需的字符设备和驱动程序文件。

本文研究分析了基于 Volta 架构的图形处理器和统一计算设备架构，如合作

组织，动态平衡度等编程模型。笔者通过在实际工作中遇到的问题和需求为

Docker 提供了运行时的扩展应用，它无缝隙的支持基于 Volta 架构的图形处理

器，并在容器内能实现所有最新 CUDA 提供的技术方案，它为容器化应用程序提

供了访问图形加速器和主机系统的专用机制，从而有效地解决了高性能计算资源

在容器内的移植性问题。

在实际工作中，优化后的容器将应用程序包装到隔离的虚拟环境中，通过此

项技术，大大简化了 CUDA 和操作系统在数据中心环境部署工作，运维和测试人

environments and take full advantage of the resources of the graphics processing

accelerator while resolving complex software stack dependencies. Containers are

typically used to seamlessly deploy CPU-based applications on multiple machines.

With this use case, the container is hardware-independent and

platform-independent. The use of NVIDIA graphics processing accelerator is clearly

not the case because it uses dedicated hardware and requires the installation of

NVIDIA drivers. As a result, the Docker Engine itself does not support encapsulating

the NVIDIA graphics accelerator inside the container.

A container is a lightweight virtualization technology that is portable, easy to build

and deploy, takes up less space and runs less time. Solve the problem of deploying

This paper analyzes and analyzes the graphics processor based on Volta

architecture and the unified computing device architecture, such as cooperative

organization, dynamic balance and other programming model. The author provides

Docker with run-time extended applications through the problems and requirements

encountered in real work. It seamlessly supports the graphics processor based on

the Volta architecture and enables the implementation of all the latest CUDA

technology solutions in the container. Provides a specialized mechanism for

accessing graphical accelerators and host systems for containerized applications,

effectively addressing the portability of high-performance computing resources in

containers.

Heterogeneous System, CUDA

目录

第一章 绪论

1.1 课题背景和意义

1.2 国内现状和国外研究进展

1.3 课题研究的创新点

1.4 论文的主要目标

1.5 本文的组织结构

第二章 系统相关技术研究

2.1 虚拟化容器技术

第三章 基于 CUDA 的高性能平台设计

3.1 CUDA 编程模型与环境

3.2 合作组织

3.3 动态平衡度

3.4 NVML

3.5 GPU P2P