High Performance Computing
Posts tagged Supercomputers
[SSC]Rmax Rpeak Nhalf
Aug 17th
2010年六月,Top500上,深圳国家计算中心的星云超级计算机(Nebulae – Dawning TC3600 Blade, Intel X5650, NVidia Tesla C2050 GPU)排名第二。Linpack测试1.27PFLOPS,是继天河1号我国又一台达到千万亿次处理能力的超级计算机,亚洲第一,世界第二。
Rmax,Rpeak,Nhalf是Linpack基准测试中的几个衡量标志。具体含义如下:
- Rmax:其是根据HPL基准测试程序,对超级计算机所能达到的最大性能的衡量指标,以GFlop/s为单位。也是Top500衡量计算机最重要最首要的指标。
- Rpeak:其是超级计算机所能达到的理论峰值。指的是一段时间内(通常为一个机器时钟周期)所能完成的浮点加法和乘法操作。如Intel Itanium 2,1.5GHz,每个时钟周期能完成4个浮点操作,换言之,其理论峰值为1.5*4=6GFlops。
- Nhalf:该值指的是当超级计算机达到其峰值速度一半时所需要的问题规模,通常可以用来很好的衡量计算机的带宽,通常为Nmax的1/10较好。
[SSC]BlueGene/L System – Node Overview
Aug 17th
一个BG/L结点ASIC,包含两个标准PowerPC 440处理核心,每个处理核心又有一个PowerPC 440 FP2核心(”Double” 64-bit Floating-Point Unit)。440是一个标准的32-bit微处理核心。
440CPU本身不通过硬件来提供SMP支持,所以两个核不是L1 Cache一致的。用一个LockBox来实现处理器到处理器间的通信一致。每个核拥有一个由Data Pre-fetch Engine控制的2KB大小的L2 Cache,核间通信的快速SRAM阵列,L3 Cache目录,4MB EDRAM构成的L3 Cache,外围DDR内存控制器,千兆以太网适配器,JTAG接口等。两个核间的L2和L3具有一致性。
More >
[SSC]BlueGene/L System – System Architecture
Aug 16th
在看Membership Service时,其中也谈到了IBM BlueGene/L。作为曾经Top500第一的超级计算机,我觉得还是挺有必要深入的了解一下。第一部分,介绍一下BlueGene/L(BGL,蓝色基因)的系统架构(System Architecture)。
BlueGene/L超级计算机,作为美国ASCI战略计划的一部分,是IBM与Lawrence Livermore National Laboratory一同开发研制的。拥有2^16=65536个结点的超大规模并行系统,基于一种全新的System-on-a-chip(SoC,片上系统)架构(Soc将之后详细介绍),其峰值处理能力到达360TFLOPS。
具有2^16=65536个计算节点的BlueGene/L系统,通过64*32*32的3D Torus网络结构构成。每个结点由一个ASIC和内存组成,支持达2GB的本地内存,并具有9个256MB内存的SDRAM-DDR内存芯片。用来构成结点的ASIC是完全基于IBM的Soc技术的。每个结点才11.1平方毫米大,具有高密度处理能力。ASIC使用的是700MHZ的IBM PowerPC 400处理器。
其组成可以简单表述如下:
- Node:2 Processors per Node
- Compute Card:2 Nodes per Compute Card
- Node Board:16 Compute Cards per Node Board
- Midplane:16 Node Boards(512-node)per Midplane
- Rack:2 Midplanes(1024-node)
- BlueGene/L:64 Racks(65536-node)
3D Microchips for More Powerful and Environmentally-Friendly Computers
Jan 12th
Years ago, home computers were powered by single-core processors, and in time, dual- and quad-core designs emerged which meant that two of four cores were placed alongside each other.
NOW, experts are looking to create 3D microprocessors, in which the cores will not be stacked next to each other, but on top of each other.
The vertical stacking allow for a much higher transfer speed between the cores – up to 10 times faster than currently possible – as well as for improved heat efficiency and overall performance.
But, efficiency is not the only reason why 3D architectures are researched. It is predicted that the supercomputers of 2100 would theoretically use up the whole of the US’ electrical supply. A new, revolutionary cooling system that would further reduce temperatures inside future 3D microprocessors is being studied now. The architecture of 3D microprocessor and its comparison with today’s multichip can be seen in the right picture.


Recent Comments