面向GPGPU系统中存储访问优化关键技术的研究

GPGPU improves the performance of data processing in a wide variety of data-intensive applications. The in-depth analysis reveals that, when GPGPU with a memory miss accesses data stored in CPU-side storage device, it takes a long path, incurs data transmission and management overhead, and then the data pipeline will be stalled by the high latency, which reduces GPGPU performance in storage-limited applications and aggravates its memory wall problem. Therefore, this project proposes an optimized scheme of storage access for GPGPU systems. 1) Based on solid state disks (a.k.a. SSDs), we will build a GPGPU memory-SSD tiered storage model, including system architecture, software stack, performance model, and etc.; 2) A set of key techniques and schemes will be proposed for the GPGPU memory-SSD tiered storage system, including the data access interface and direct path-based programming model, spatial detection and transmission management-based runtime system, data distribution and access consistency-based data storage management, and etc.; 3) We will design a prototype and measurement methods for the novel GPGPU memory-SSD tiered storage system. Through the implementation of this projected, our goal is to narrow down data-path length from GPGPU to SSDs, decrease the latency, reduce overload in software stack, shorten pipeline stall time, address the GPU memory-wall challenge, and improve the GPGPU system performance. This project aims to investigate storage access-based optimization for GPGPU systems featured with high performance and large capacity. And its prospective research results will not only deepen improvement of GPGPU performance, but also provide valuable technical cases for the performance optimization of heterogeneous systems for big data processing.

GPGPU提高了数据密集型应用处理的性能。分析发现，存储受限型应用中，当GPGPU访存失效且需访问CPU端存储空间时，其访问路径较长，软件栈内数据传输和管理开销较大，造成访问延时较高，导致流水线阻塞，加重了GPGPU内存墙问题并降低其性能。本项目拟提出一种面向GPGPU系统的存储访问优化方案。1）拟构建一种基于固态盘的GPGPU内存-SSD分层存储系统模型（含架构、软件栈及性能模型等）；2）拟提出一组关键技术（含基于编程接口和数据访问通道的编程模型、基于空间检测和数据传输管理的运行时系统、基于数据分布和存储访问一致性的数据存储管理等）；3）拟设计一种系统原型及测试方案。达到缓解GPGPU内存墙问题并提高其性能的目的。本项目旨在研究一种高性能、大容量的面向GPGPU系统的存储访问优化方案，其预期成果将丰富GPGPU系统性能优化的研究，为面向大数据处理的异构系统性能优化提供方案借鉴和技术参考。

针对分层存储架构，构建跨层存储管理软件层，实现任务从CPU端卸载到设备端，基于此系统，提出了一种以计算力与时间比值为参考值的任务卸载和调度策略，这对于实现任务卸载到跨层的任务交互具有重要的意义。以提高固态盘内部数据处理性能为目标，结合当前固态盘所具有的大容量的特色，我们针对固态盘内部软件栈的设计，是跨层存储系统的设备端方面，提出高效日志管理和缓存管理策略，这对缓解固态盘现有性能访问抖动性，对系统整体性能的优化具有重要的作用。实现系统级通信协议、数据传输通路及任务卸载框架开展研究，基于语义管理模块，实现跨层主机端与设备端之间的任务卸载，跨层软件栈直接按的数据传输由特定的传输协议保证，并采用同步的方式保证跨层存储软件栈之间的可靠数据传输。以键值存储系统为对象，降低数据传输量，充分利用近数据计算模型实现对任务卸载，在设备端执行任务，实现主机端和设备端二者协同处理计算任务的目的。针对新型闪存固态盘的缓存管理策略。首先，该缓存策略通过主动写回缓存中的脏数据解决了在密集写负载的工作环境下，基于CFLRU缓存管理策略的一系列缓存管理策略没有足够的干净数据进行替换的问题，使得缓存替换的数据也接近100%是干净数据，极大减少了缓存替换的开销。缓存策略利用闪存内部多余的芯片级并行性，将缓存中的数据并行地写入闪存芯片中，并没有增加额外的带宽负载。同时缓存策略通过冷热分离，干净数据阈值等方案解决了过去提前写回缓存策略造成的写闪存操作过多，极大影响闪存寿命的问题。