Intel xeon phi – HP Insight Cluster Management Utility User Manual

Page 151

Advertising
background image

A network bridge must be configured with the IP address of the local compute node (/etc/
sysconfig/network-scripts/ifcfg-br0

).

The network device must be associated with the bridge (/etc/sysconfig/
network-scripts/ifcfg-ethX

).

Each host-side Xeon Phi card Ethernet device must be associated with the bridge (/etc/
sysconfig/network-scripts/ifcfg-micX

).

Each Xeon Phi master configuration file must be configured with the Xeon Phi host name, IP
address, and network configuration (/etc/sysconfig/mic/micX.conf).

Each Xeon Phi card host name must be configured in the Xeon Phi file system (/opt/intel/
mic/filesystem/micX/etc/sysconfig/hostname

).

Each MIC-side Xeon Phi card Ethernet device must be configured with the Xeon Phi IP address
(/opt/intel/mic/filesystem/micX/etc/sysconfig/network/ifcfg-mic0).

Sync the root ssh keys in each Xeon Phi card file system.

The settings listed above must be scripted to configure correctly on each compute node, except for
the initial /etc/hosts configuration, which must be done by the user beforehand. However, HP
Insight CMU must be scripted to copy the host compute node /etc/hosts file into each Xeon
Phi file system.

8.7.1 Intel Xeon Phi card IP address and host name assignment algorithm

HP Insight CMU cloning and diskless deployment both configure the per-node host name and IP
addresses for the cluster management network. When Intel Xeon Phi cards are present on each
compute node and configured with independent host names and IP addresses, HP Insight CMU
needs a way to determine the host name and IP address that will be assigned to each Intel Xeon
Phi card on each compute node. There are several ways to accomplish this. The recommended
way, which is employed in the examples below, is to assign <compute_node_hostname>-mic0
as the host name of the first Intel Xeon card on each compute node, and
<compute_node_hostname>-mic1 as the host name of the second Intel Xeon Phi card on each
compute node, and so on. These examples only focus on two Intel Xeon Phi cards per compute
node, but this algorithm scales to any count.

Now that the Intel Xeon Phi card host name can be determined from the local compute node host
name, you can easily search the /etc/hosts file (using grep) for the corresponding IP address
assigned to this Intel Xeon Phi card. With this method, we can programmatically determine the
host name and IP address of any Intel Xeon Phi card on any compute node.

Note that the success of this algorithm depends on having the Intel Xeon Phi card host name and
IP address pre-configured in the /etc/hosts file on the HP Insight CMU management node. This
must be done before any of the following scripting examples can be employed. The benefits of
this requirement are:

The hostname-to-IP address mapping definition in /etc/hosts ensures that any Intel Xeon
Phi card host name can be resolved to its IP address from any node or Xeon Phi card within
the cluster.

The following cloning or diskless provisioning examples will be successful in deploying an
OS and configuring the Intel Xeon Phi cards with unique host names and IP addresses.

IMPORTANT:

You must configure the host name (in the format recommended above) and IP

address of your Intel Xeon Phi cards in /etc/hosts on the HP Insight CMU management node
before attempting any cloning or diskless image building.

8.7 Support for Intel Xeon Phi cards

151

Advertising