HPCS supports five different cluster topologies. Each topology
has implications for performance and accessibility. The topologies
involve up to three different networks: public, private, and MPI.
In most cases, NAT (using RRAS) is automatically installed on the
head node.
Figure 2-1 shows a head node
on a public and private network with compute nodes on the private
network only. This topology is referred to as Two-Network Topology
and is supported by HP Cluster Platform and HP Cluster Platform Express.
See Figure 2-6.
Figure 2-2 shows all nodes on
a public and private network.
Figure 2-3 shows the head node
on a public and private network with the compute nodes on the private
network only with a second private network for MPI. This topology
is referred to as Three-Network Topology and is supported by HP Cluster
Platform and HP Cluster Platform Express. See Figure 2-7.
Figure 2-4 shows all nodes on
a public and private network with a second private network for MPI.
Figure 2-5 shows all nodes on
a public network only.
Only two of the five topologies previously shown map to HP Cluster
Platform topologies. Figure 2-6 shows the simplest topology, based on in-band (shared) use of the HP ProCurve switch
as the MPI fabric. The same network provides routing for cluster administrative
and job management traffic. A second network is provided through the
site LAN connection.
The following cluster features are shown in Figure 2-6:
| 1 | Site LAN connection |
| 2 | LAN network, connected to a network interface card (NIC) installed
(or embedded) in the head node |
| 3 | The server designated as the head node (control node in HP Cluster
Platform terminology) |
| 4 | NIC dedicated to the LAN network |
| 5 | Connection to the server management interface, a dedicated hardware
interface that listens for, and executes commands received through
the Ethernet network. This card is also known as the management processor
(MP), or integrated lights-out (iLO) card, depending on the model
of server. Where supported by the operating environment, this connection
is used for server hardware management functions, such as power-off
or boot. |
| 6 | NIC dedicated to the Message Passing Interface (MPI) |
| 7 | Compute nodes |
| 8 | Common Gigabit Ethernet network, used both for MPI traffic,
and for in-band job management, and cluster administrative functions |
| 9 | Gigabit Ethernet switch, used as the system interconnect |
Figure 2-7 shows a
topology based on a dedicated MPI fabric. This might be built around Gigabit Ethernet or InfiniBand.
A second Ethernet network provides routing for cluster administrative,
and job management traffic. A third network is provided through the
site LAN connection.
The following cluster features are shown in Figure 2-7:
| 1 | Site LAN connection |
| 2 | LAN network, connected to a network interface card (NIC) installed
(or embedded) in the head node |
| 3 | The
server designated as the head node (control node in HP Cluster Platform
terminology) |
| 4 | NIC dedicated to the LAN network |
| 5 | NIC connected to a
dedicated Gigabit Ethernet MPI network, if present |
| 6 | Connection to the server management interface, a dedicated
hardware interface that listens for, and executes commands received
through the Ethernet network. This card is also known as the management
processor (MP), or integrated lights-out (iLO) card, depending on
the model of server. Where supported by the operating environment,
this connection is used for server hardware management functions,
such as power-off or boot. |
| 7 | NIC connected to a shared Gigabit
Ethernet MPI network, if present |
| 8 | Host Channel Adapter (HCA)
card, connected to an InfiniBand MPI network, if present |
| 9 | Compute
nodes |
| 10 | Ethernet network for cluster administration |
| 11 | Gigabit Ethernet MPI network, if present |
| 12 | InfiniBand MPI network, if present |
| 13 | Gigabit Ethernet system interconnect,
if present |
| 14 | Gigabit Ethernet switch for cluster administrative network,
if present |
| 15 | InfiniBand system interconnect, if present |