Silicon Graphics Prism, Руководство пользователя

Страница: 84 / 250

62 007-4701-003
3: Configurations and Cabling
NUMAlink Cabling in Ring Topologies
The basic Silicon Graphics Prism uses a type of network configuration that is referred to
as a ring topology. As the name implies, the network connection between the base
compute module, graphics module, and optional CPU expansion or CMPX modules
forms a ring. A message is passed around the ring until it reaches its destination. The data
flow of this ring topology flows in both directions, enabling the modules to have direct
connection to two other modules and providing an alternative path if a connection fails
between two modules.
The Silicon Graphics Prism system may also be configured in routed configurations,
which are described in “NUMAlink Cabling in Routed Systems” on page 65.
The bisection bandwidth per node depends on the number of modules on the ring, as
well as what type of modules they are; the bisection bandwidth per node is greatest when
there are only two modules on the ring. See Table 3-3.
Table 3-3 Bisection Bandwidth of Ring Topology
Module Count Bisection Bandwidth
2 3.2 GB/s/node
3 (with 2 pipes) 2.13 GB/s/node
3 (with 4 pipes) 1.83 GB/s/node
4 (with 2 pipes) 2.13 GB/s/node
4 (with 4 pipes) 1.6 GB/s/node