Lustre discuss - Oct 2008 - lustre/drbd/heartbeat setup [was: drbd async mode]

Hello,

at last a first version of our setup scenario is ready.

Please consider this as a general guideline. It may contain errors.
We know that some things are done differently in the lustre community i.e. 
placing MDS and MDT on seperate machines.

Please let me know if you find bugs or if things can be improved.

"There is more than one way."

Regards
Heiko
-------------- next part --------------
A non-text attachment was scrubbed...
Name: lustre_drbd_hearbeat_setup.pdf
Type: application/pdf
Size: 421143 bytes
Desc: not available
Url :
http://lists.lustre.org/pipermail/lustre-discuss/attachments/20081002/1651c518/attachment-0001.pdf

Heiko Schroeter wrote:
> Hello,
>
> at last a first version of our setup scenario is ready.
>
> Please consider this as a general guideline. It may contain errors.
> We know that some things are done differently in the lustre community i.e. 
> placing MDS and MDT on seperate machines.
>
> Please let me know if you find bugs or if things can be improved.
>
> "There is more than one way."
>   
With drbd 0.7 I had no problems with MDS but 8.x series were extremly 
slow to me. I use the same scenario as you: sw raid, but I don''t know, 
if it matters at all.

I get this help, still I haven''t tried:

"look up the no-disk-flushes and no-md-flushes config options.

don''t use on io-subsystems with volatile caches, though."



tamas

Hi,

Thanks for that.I was thinking about trying drbd on my MDSs so I find
your PDF very useful.

Heiko Schroeter wrote:

Hello,

at last a first version of our setup scenario is ready.

Please consider this as a general guideline. It may contain errors.
We know that some things are done differently in the lustre community i.e. 
placing MDS and MDT on seperate machines.

I guess you meant placing MGS and MDT  on separate machine? I know that
lustre has lots of abbreviation''s and it is important no to mix them to
avoid confusing others.

Please let me know if you find bugs or if things can be improved.

"There is more than one way."

Regards
Heiko

Regards,

Wojciech

_______________________________________________
Lustre-discuss mailing list
Lustre-discuss-aLEFhgZF4x6X6Mz3xDxJMA@public.gmane.org
http://lists.lustre.org/mailman/listinfo/lustre-discuss


_______________________________________________
Lustre-discuss mailing list
Lustre-discuss-aLEFhgZF4x6X6Mz3xDxJMA@public.gmane.org
http://lists.lustre.org/mailman/listinfo/lustre-discuss

Am Donnerstag, 2. Oktober 2008 10:32:44 schrieb Wojciech
Turek:
>  Hi,
>
>  Thanks for that.I was thinking about trying drbd on my MDSs so I find your
> PDF very useful. Heiko Schroeter wrote:
> Hello,
>
> at last a first version of our setup scenario is ready.
>
> Please consider this as a general guideline. It may contain errors.
> We know that some things are done differently in the lustre community i.e.
> placing MDS and MDT on seperate machines.
>
>  I guess you meant placing MGS and MDT? on separate machine? I know that
> lustre has lots of abbreviation''s and it is important no to mix
them to
> avoid confusing others.
Yes, sorry for that.
>
> Please let me know if you find bugs or if things can be improved.
>
> "There is more than one way."
>
> Regards
> Heiko
>
>
>  Regards,
>
>  Wojciech
>
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss

Hi,

read your instructions - that''s pretty much the setup we are using,
too.
And it works very well, drbd 0.8 non-withstanding, but on a hardware raid.
I do not quite understand your remark about not using an extra net for 
drbd - have you tried putting the name that''s in your drbd.conf
together
with the other IP into /etc/hosts ? My guess is that the performance of 
your MDS-pair is influenced by drbd doing its job - I would keep that 
separate from the Lustre data stream.
The machines we are planning to use in our next cluster are actually 
equipped with four network interfaces - two (bonded) for Lustre, one for 
drbd and one for heartbeat - those serial cables only give me error 
messages and headaches.

We have separate partitions for MGS and MDT - on one machine. Didn''t 
understand that this would not be the Lustre way? This way at least one 
doesn''t have to worry about a super fast connection between MGS and
MDT.

Is there a particular reason for not managing the IP via heartbeat? At 
least it''s easier to setup than the drbddisk and Filesystem resources.

Regards,
Thomas

Heiko Schroeter wrote:
> Hello,
> 
> at last a first version of our setup scenario is ready.
> 
> Please consider this as a general guideline. It may contain errors.
> We know that some things are done differently in the lustre community i.e. 
> placing MDS and MDT on seperate machines.
> 
> Please let me know if you find bugs or if things can be improved.
> 
> "There is more than one way."
> 
> Regards
> Heiko
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss