Q-Logic IB6054601-00 D 用户手册

2 – InfiniPath Cluster Administration

Performance and Management Tips

IB6054601-00 D

2-17

Q

0zwxSL7GP1nEyFk9wAxCrXv3xPKxQaezQKs+KL95FouJvJ4qrSxxHdd1NYNR0D

avEBVQgCaspgWvWQ8cL 

0aUQmTbggLrtD9zETVU5PCgRlQL6I3Y5sCCHuO7/UvTH9nneCg==

Change the file to mode 600 when finished editing.

4. On each node, the system file 

/etc/ssh/sshd_config

 must be edited, so that 

the following four lines uncommented (no 

#

 at the start of the line) and are set 

to 

yes

. Each of these lines is normally present, but commented out and set to 

no

 by default.

RhostsAuthentication yes

RhostsRSAAuthentication yes

HostbasedAuthentication yes

PAMAuthenticationViaKbdInt yes

5. After creating or editing these three files in steps 2, 3 and 4, 

sshd

 must be 

restarted on each system. If you are already logged in via 

ssh

 (or any other 

user is logged in via 

ssh

), their sessions or programs will be terminated, so do 

this only on idle nodes. Tell 

sshd

 to use the new configuration files by typing 

(as root):

# killall -HUP sshd

This will terminate all ssh sessions into that system. Run from the 
console, or have a way to log into the console in case of any problem.

At this point, any user should be able to login to the 

ip-fe

 front end system, and 

then use 

ssh

 to login to any InfiniPath node without being prompted for a password 

or pass phrase.

MPI jobs that use more than 8 processes per node may encounter an SSH throttling 
mechanism that limits the amount of concurrent per-node connections to 10. If you 
need to use more processes, you or your system administrator should increase the 
value of ’MaxStartups’ in your sshd configurations. See 

 for an 

example of an error message associated with this limitation.

The following section gives some suggestions for improving performance and 
simplifying management of the cluster.

An important step that the cluster administrator can take to enhance application 
performance is to minimize the set of system services running on the compute