The manual says that Lustre assigns an operational thread to each logical
interface, so if you''re using bonding, you''ll get a single
transaction
thread for your bond, and not one-per physical interface. There are
presumably both advantages and disadvantages to this approach, but I
haven''t
yet had a chance to test it in the field.
As far as configuration, you need to specify which interfaces in your
modprobe.conf lines, something like this, i.e.
-- cut --
options lnet networks=tcp0(bond0),tcp1(eth3)
-- cut --
And then in your mount statements, specify which lnet interface you wish to
use to communicate, i.e.
-- cut --
mount -t lustre mds0 at tcp0:/lustre
mount -t lustre mds1 at tcp1:/lustre
-- cut --
hth,
Klaus
On 10/23/07 12:16 PM, "Brock Palen" <brockp at umich.edu>did
etch on stone
tablets:
> Hello,
> In reading the manual about bonding (we will need to bond our Gige)
> I get the impression that we should let lustre take care of it and
> not use the linux bonding drivers. Am I correct in this assumption?
> If so I would have 2 interfaces each with their own IP, how does the
> client know to use both interfaces? Is this just advertised by the
> MGS to the client when mounted? Or does a client choose a random
> interface and does all IO to that OSS though that interface for that
> transaction?
>
> Also our OSS will have 2 separate devices, I would like to be
> separate stripe targets for lustre. (ie if using 1 OSS lfs
> getstripe, would show a large files striped over 2 OST''s) Is this
> fine? or should we software raid them together and make one large
> OST on top?
>
> Thank you
>
> Brock Palen
> Center for Advanced Computing
> brockp at umich.edu
> (734)936-1985
>
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at clusterfs.com
> https://mail.clusterfs.com/mailman/listinfo/lustre-discuss