Steve vmwx at 03:36, 22 February 2006

2006-02-22T03:36:54Z

Steve vmwx at 03:35, 22 February 2006

2006-02-22T03:35:22Z

New page

As an example, say you run your nodes with a FastEther port for the admin and reporting functions and leave the GigaEther for the MPI traffic.

IP addresses are setup accordingly and they have their own host names. eg. borg1 is the GigaEther (eth1) while borg1_m (eth0) is the "management" interface. Now, if you don't tell MPI that there's a difference then it's going reach out and touch the machines with eth0.

To limit the MPI traffic to a specific interface you need to tell it which interface to use. A bit if string work in your mpirun.args file does the trick.

In this case:

MPI_HOST=`hostname | sed "s/_m//"`

This slices the "_m" off the end of the hostname.

(thanks to Reuti at the Uni of Marburg, Germany from the Beowulf email list)

← Older revision		Revision as of 03:36, 22 February 2006
Line 1:		Line 1:
−	As an example, say you run your nodes with a FastEther port for the admin and reporting functions ~~and~~ leave the GigaEther for the MPI traffic.	+	As an example, say you want to run your nodes with a FastEther port for the admin and reporting functions but leave the GigaEther for the MPI traffic.

	IP addresses are setup accordingly and they have their own host names. eg. borg1 is the GigaEther (eth1) while borg1_m (eth0) is the "management" interface. Now, if you don't tell MPI that there's a difference then it's going reach out and touch the machines with eth0.		IP addresses are setup accordingly and they have their own host names. eg. borg1 is the GigaEther (eth1) while borg1_m (eth0) is the "management" interface. Now, if you don't tell MPI that there's a difference then it's going reach out and touch the machines with eth0.

Limiting MPI (1.2x) traffic to one NIC - Revision history

Steve vmwx at 03:36, 22 February 2006

Steve vmwx at 03:35, 22 February 2006