美文网首页分布式数据计算
4. 高性能计算机的历史概览

4. 高性能计算机的历史概览

作者: 喜欢苹果的梨 | 来源:发表于2018-12-23 02:32 被阅读19次

    都知道,“工欲善其事必先利其器”。

    为了完成前面以天气预报那样的计算,就需要巨大的计算能力!

    计算机的早期历史是应该去了解的 - 3个人和两个模型(图灵机,冯诺依曼架构)

    不知道的,就请移步去查查吧:) 

    1962年的并行 - IO和CPU间的并行 - Channel 

    按照“Grepory R. Andrews. Foundations of Multithreaded, Parallel and Distributed Programming”书中所述,“Concurrent programming oringinated in 1962 with the invention of channels which are independent device controllers that make it possible to have a CPU execute a new application program at the same time that I/O operations are being executed on behalf of other, suspendended application programs.”

    之后没多久就有了多处理器的系统 - “In the late 1960s, hardware designers developed multiple processor machines”.

    The LARC was the first true supercomputer – it was designed for multiprocessing, with 2 CPUS and a separate I/O processor.

    1966 年Flynn's Taxonomy很有前瞻地给出了高性能系统的分类,但是,MIMD太粗糙了

    1970s开始的Networking 引发了所谓的分布式系统

    还是“Grepory R. Andrews. Foundations of Multithreaded, Parallel and Distributed Programming”书中所述:“Computer networks were introduced in the late 1970s and early 1980s. The Arpanet supported wide-area computing, and the Ethernet established local-area networks. Networks gave rise to distributed programming, which was a major topic of the 1980s and became even more important in the 1990s. ”

    The Vector Years – 1974 :The CDC Star-100 was one of the first machines to use a vector processor

    - Used “deep” pipelines (25 vs. 8 on 7600) which need to be "filled" with data constantly and had high setup cost. The vector pipeline only broke even with >50 data points in each set.

    - But the number of algorithms that can be effectively vectorised is very low and need careful coding otherwise the high vector setup cost dominates.

    - And the basic scalar performance had been sacrificed in order to improve vector performance => machine was generally considered a failure.

    Vector processors have high-level operations that work on linear arrays of numbers: "vectors"

    Cray 系列是Vector machines 的代表

    Cray-1

    Innovation implies in the history - GPU:跟Vector processing 的原理很像的

    GPU (graphics processing unit ) is a RISC specialized processor that offloads 3D graphics rendering from microprocessor.

    - 1970s: ANTIC and CTIA chips provided for hardware control of mixed graphics and text modes, sprite positioning and display, and other operations based on Atari 8-bit computers.

    - 1980s: IBM Professional Graphics Controller was one of the very first 2D/3D graphics accelerators available for the IBM PC, released in 1984. But it was expensive, slow and lack of compatibility.

    也有了一些有趣的应用架构 - Grid 和 P2P等

    Cluster 起来了 - 是Google起家的架构,也就成了现在的流行架构呢

    NASA Beowulf Project – 1994

    Cluster consists of:

    - Nodes(master+computing)

    - Network

    - OS

    - Cluster middleware: Middleware such as MPI which permits compute clustering programs to be portable to a wide variety of clusters

    其实,从硬件的角度看,上面的von Neumann, Multiple Processing machine (Vector, GPU)和Cluster 就是HPC计算机的基本元素 - 实际系统的建立也就是那些基本元素的组合吧

    架构的小节

    峰回路转,Cloud概念出来了 - 不过, 可以理解为将前面的那些概念虚拟化

    https://www.researchgate.net/figure/Evolution-of-Various-Computing-Models_fig1_275405572

    相关文章

      网友评论

        本文标题:4. 高性能计算机的历史概览

        本文链接:https://www.haomeiwen.com/subject/gikgkqtx.html