Lustre discuss - Dec 2006 - Manually load config XML every time?!!!!

Hello all,
   
  I''m new to lustre and was hoping someone could enlighten me to a few
questions. We''re running lustre version lustre.1.4.7.1 on centOS kernel
2.6.9-42. We currently have 4 OSS''s, one MDS and 11 compute nodes.
  We''ve had quite a log of issues lately which required rebooting
everything. Lustre currently is set to be started manually and not as a service
due to dependency issues we''ve faced. So I guess my first question is:
  -What is the proper order in which this type of config should booted? I should
also mention that we also have a "head node" which I believe is no
more than a client itself.
   
  When the systems are up, I start the services for lustre and the status shows
that lustre is running, but I still have to tell lustre where my xml file is so
it loads (lconf --node client /lustre/config.xml). Why?
   
  And finally, we''ve had plenty of I/O errors which (storage
interconnect vi infiniband over IP) we believe may be due to timeout issues.
What are the "best practices" for setting ldlm timeouts and timeouts
for such a config? and can I set those permanently in the configuration rather
than having to use sysctl, maybe through lcm or lctl?
   
  Thanks a bunch!

 
---------------------------------
Everyone is raving about the all-new Yahoo! Mail beta.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://mail.clusterfs.com/pipermail/lustre-discuss/attachments/20061201/6a341f2e/attachment-0001.html

On Fri, 1 Dec 2006, Syed Haider wrote:
>  -What is the proper order in which this type of config should booted? I 
> should also mention that we also have a "head node" which I
believe is
> no more than a client itself.
Start the MDS first, then the OSSes.
>  When the systems are up, I start the services for lustre and the status 
> shows that lustre is running, but I still have to tell lustre where my 
> xml file is so it loads (lconf --node client /lustre/config.xml). Why?
AFAICS the service script expects the XML to be /etc/lustre/config.xml, 
unless LUSTRE_CONFIG_XML is set in /etc/sysconfig/lustre.
>  And finally, we''ve had plenty of I/O errors which (storage
interconnect
> vi infiniband over IP) we believe may be due to timeout issues. What are 
> the "best practices" for setting ldlm timeouts and timeouts for
such a
> config? and can I set those permanently in the configuration rather than 
> having to use sysctl, maybe through lcm or lctl?
Your distro probably has /etc/sysctl.conf for this purpose.

Bull has also released Lustre administration tools, which are useful to 
set up and start Lustre systems from a single admin node. Not sure the 
current (and old) released patch applies cleanly to 1.4.7, but you could 
try. See: http://www.bullopensource.org/lustre/


-- 
Jean-Marc Saffroy - jean-marc.saffroy@ext.bull.net

On Fri, 1 Dec 2006, Jean-Marc Saffroy wrote:
> On Fri, 1 Dec 2006, Syed Haider wrote:
>
>>  -What is the proper order in which this type of config should booted?
I
>> should also mention that we also have a "head node" which I
believe is no
>> more than a client itself.
>
> Start the MDS first, then the OSSes.
Actually it''s the OSSes first, sorry.


-- 
Jean-Marc Saffroy - jean-marc.saffroy@ext.bull.net