精华公布栏

发信人: carl.bbs@bbs.swjtu.edu.cn (师哥), 信区: cnunix
标  题: 曙光1000 知识讲座（八）
发信站: 锦城驿站 (Sat Dec  6 20:53:17 1997)
转信站: Lilac!ustcnews!ustcnews!sjtunews!swjtu!swjtubbs
出  处: xnjd21.swjtu.edu.cn

曙光1000性能评介（2）

  The Dawning-1000 came up to the advanced world levels of the early 1990s.

***   The self-developed wormhole routing chip creatively exploited the
    combination of asynchronous mode and synchrounous mode, and it is one of
    the several most advanced routing chips in the world.

***    The parallel optimizing compiler of the Dawning-1000 is assessed as one of
    the few most excellent systems in the world by the world famous colleague
    scholars like Arvind in MIT and Gao Guang-Rong in the University of McGill.
    The Paravision, one of the tools on the Dawning-1000, is very advanced in
    the world, on which the person in charge of the technologies for the
    Japanese Fifth Generation Machina Project spend many years but failed.

***    The Dawning-1000 exploits the advanced architecture. Its operating systems,
    software environments and tools, and user interface all use the
    international standards. The overall level of its software is equivalent to
    that of cuerent MPP systems.

***    Testings have shown that the performance of the Dawning-1000 is equal to
    that of famous MPP systems of the same class and scale, and on some
    aspects, eg., the largest problem size it can solve, it exceeds the system
    of the same class and scale.

        The table below gives out the highest performance comparation between
the Dawning-1000 and some most advanced MPP systems in the world on Highly
Parallelu Computing Testing. The figures of machines in the table are extracted
from J.Dongarra's technical report "Performance of Various Computers Using
Standard Linear Equations Software" of April 13, 1995. We can see from the
table that the highest performance of the 32-node Dawning-1000 (40 MHZ) is
comparable to that of the Paragon, and superer to that of Delta. The 32-node
Dawning-1000 can solve the linear equations of order 15,000, but the 128-node
Paragon can only solve the linear equations of 12,000. The speed for solving
the linear equations of order 1000 on a single processor of the Paragon (50
MHZ) is 34 MFLOPS, but that on the Dawning-1000 is 35.4 MFLOPS.

        Table of the Highest Performance for Sloving Linear Equations on MPPs

Machine Name     Number of          R(max)     N(max)     N(1/2)
                 Computing Nodes    GFLOPS     Order      Order

Dawning-1000     32                 1.12       15000      4000

Paragon XPS      256                7.6        16000      4000
(50 MHZ)         128                4.0        12000      3000
                 64                 2.0        8000       2000

Delta            192                5.2        15000      4500
(40 MHZ)         128                3.5        12500      3500
                 64                 1.7        8000       2500
                 32                 0.9        6000       2000

IBM SP-2         32                 6.1        19000      4088
                 16                 3.0        13500       -

CM-5             128                7.7        18432      8192
                 64                 3.8        13056      6016

Cray T3D32       32                 3.2        14592      1616

SGI Challenge    36                 1.28       8000       2000

Ncube 2          256                0.482      10784      1504

Maspar MP-1266   16384              0.47       11264      1280

[Notes] R(max): the highest speed that the machine can reach
        N(max): the order of linear equations when speed reaches R(max)
        N(1/2): the order of linear equations when speed reaches 1/2 R(max)

--

  三人同行   必有我师
       倘佯校网    愿交吾友
          有朋自网上来,幸哉、喜哉...
          永远的朋友Carl

※ 来源:．锦城驿站 bbs.swjtu.edu.cn．[FROM: 202.115.65.79]