发信人: carl.bbs@bbs.swjtu.edu.cn (师哥), 信区: cnunix
标 题: 曙光1000 知识讲座 (四)
发信站: 锦城驿站 (Sat Dec 6 20:41:57 1997)
转信站: Lilac!ustcnews!ustcnews!sjtunews!swjtu!swjtubbs
出 处: xnjd21.swjtu.edu.cn
曙光1000硬件 (1)
3. Hardware of the Dawning-1000
3.1 System Architecture
The Dawning-1000 is a loosely-coupled message passing massively
parallel processing system. It is made up of 32 computing nodes (each of which
is based on the i860 microprocessor), the system service subsystem which
contains one or two nodes, and the external storage subsystem which contains
two or three I/O nodes. All these nodes are connected by a 6x6 2-dimention
mesh. Each point on the mesh has a wormhole router. There is a system monitor
which charges the system initialization and the hardware failure diagnosis.
The working platforms for the users are SUN Sparc-II workstations (or other
workstations, eg. SGI) and some terminals. The system servic nodes, monitor,
the Sparc-II workstations and the terminals are connected by a 10 Mb/s
Ethernet. Fig 1 shows the Dawning-1000 system architecture.
| | | | | | Wormhole Routing Chip
--o---o---o---o---o---o-- ------------------> | |
| | | | | | |-y |
--o---o---o---o---o---o-- +----------+
| | | | | | +x -->| WRC |-->
--o---o---o---o---o---o-- <--| |<-- -x
| | | | | | / \---------+
--o---o---o---o---o---o-- / / | |
| | | | | | nodes/ | |+y
--o---o---o---o---o---o-- ----------------+
| | | | | | | to computing nodes
--o---o---o---o---o---o-- --------+ | 1 2 32
| | | | | | | | | | | RS232
| | | | | | |
| | | | | | |
service node computing--RS232 I/O service node monitor
(host) node node (host)
| Ethernet
|------|--------------|
| | |
SUN SGI IBM
Front-End Computers
Fig 1: the Dawning-1000 system architecture
3.2 Mesh Network
The mesh network in Fig 1 is extensible. Each node (including system
service node and I/O node) is connected with a wormhole routing chip (WRC).
The numbering of the nodes (the physical addresses of the nodes) is made up of
two parts: the numbering in X direction and the numbering in Y direction. The
relative address between two nodes is contained in messages. A WRC has 10
input/output channels. Each channel contains 8 datalines, a Tail signal which
indicates the end of a message, a sending request line (REQ) and an
acknowledging line (ACK). When a message is sent from a source node to a
destination node, it only passes through the WRCs on the way, and doesn't go
into the node processors connected by the WRCs. The mesh used X-Y routing
algorithm. Several channels can share a unidirectional physical channel.
3.3 Computing Nodes
A computing node is a plug-in board, the kernel of which is a i860xr
chip, the frequency of the chip is 40 MHZ, the word length is 64 bits, the
floating point computing speed is 80 MFLOPS for single precision and 60 MFLOPS
for double precision, the scalar computing speed is 40 MIPS. It has 32 MB DRAM
with parity check, 8 KB EPROM, and 2 KB NVRAM which is powered by batteries.
The node-mesh interface module contains two 16 KB FIFOs. The communication
bandwidth between the mesh and each node is 33 MB/s, with CRC check of message
data. The RS232 module on each node board is directly connected to a multi-user
card on the monitor in order that the monitor can supervise and disgnose the
node processor. All computing nodes run operating system and user programs.
--
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]
Powered by KBS BBS 2.0 (http://dev.kcn.cn)
页面执行时间:3.284毫秒