A High Throughput Massive I/O Storage Hierarchy for PETA-scale High-end Architectures

现代高端计算机（HEC）体系结构（由数万个或更多的处理器组成）拉大了计算能力与能够支持和维持此类计算的存储和I/O性能之间的性能差距。随着千万亿次计算（每秒1015次浮点运算）即将到来，差距将更加明显。本研究旨在解决当前I/O存储系统的有限带宽和性能问题。其主要思想是用一个“内存池”——一个由固态（闪存）存储器组成的大规模网格——取代传统的旋转磁盘，并引入一种基于改进的访问时间和带宽的新的内存层次模型。跨计算引擎嵌入和分布的内存池利用了固态存储技术的快速发展。感知I/ o的多线程程序执行模型充分利用了本地内存池的巨大带宽，同时通过主动分布式预取和软件控制的内存缓存容忍固态存储访问延迟。特别是，对不同模式的固态存储通信请求（阻塞和非阻塞，以及各种块大小）的响应，模拟固态存储模块故障，扩展共享内存响应和管理进行了检查。通过最常见的第三方I/O基准测试可以观察到该方法的有效性。本研究的智力价值可以概括为：1)开发了一种新颖的I/O体系结构模型。用于高端的千万亿级计算系统。2)相应的RAS模型的开发3)对新的I/O架构和软件模型(上述(1)和(2)进行实验研究。

Modern high-end computer (HEC) architectures (comprised of tens-of-thousands of processors or more) have widened the performance gap between the computing power and the storage and I/O performance that can support and sustain such calculations. With the petascale computing (1015 floating point operations per second) just over the horizon, the gap will be even more pronounced. This research is aimed at addressing both the limited bandwidth and the performance of today's I/O storage systems. The main idea is to replace traditional rotary disks with a 'memory pool' - a massive grid of solid-state (flash) memories, and to introduce a new memory hierarchy model based on the improved access time and bandwidth of such grid. A memory pool embedded and distributed across the compute engine leverages the rapid progression of solid-state storage technology.The I/O-aware multithreaded program execution model takes full advantage of the tremendous bandwidth to the local memory pool, while tolerating the solid state storage access latency through proactive distributed pre-fetching and software-controlled memory caching. In particular, the responses to different patterns of solid-state storage communication requests (blocking and non-blocking, and of various block sizes), simulated solid-state storage module failures, extended shared memory responsiveness and management are examined. The effectiveness of the approach is observed through the most common set of third party I/O benchmarks.The intellectual merit of this research can be summarized as:1) Development of a novel I/O architecture model. for a class of high-end petascale computing systems. 2) Development of a corresponding RAS model3) An experimental study on the new I/O architecture and software model (in (1) and (2) above).