[Lustre-discuss] InfiniBand QoS with Lustre ko2iblnd.

Isaac Huang He.Huang at Sun.COM
Tue Jun 30 23:07:33 PDT 2009


On Fri, Jun 26, 2009 at 01:42:53PM +0200, S?bastien Buisson wrote:
>
> Isaac Huang a ?crit :
>> On Wed, Jun 24, 2009 at 09:46:19AM +0200, S?bastien Buisson wrote:
>>> ......
>>> The peer's port information could be stored in the kib_peer_t 
>>> structure.  That way, it would be possible to make clients connect to 
>>> servers which  listen on different ports.
>>> What do you think?
>>
>> At this point it can't be done. But we have in our development plans
>> to implement dynamic LNet configuration which includes per-NI options
>> (i.e. it'd be possible to specify the 'service' option on a per-NI
>> basis instead of being just LND global), and once it's implemented
>> you'd be able to specify different 'service' option if you'd create
>> two server networks for the two FS.
>
> OK, if I understand correctly, the major hurdle with what I proposed is  
> that LNET is not able to get configuration information dynamically at  
> the moment, right?

Yes.

> I agree with you, I think the per-NI options in LNET would do the trick.  
> Do you have plans about when this feature would be available? Have you  
> already begun to work on it?

It's too early to make any realistic estimate at the moment. Though It's
already on the lnet roadmap, I'm not sure when we're going to start
working on it.

> If you have some pre-alpha work, we would be glad to evaluate it.

Thanks, I'll remember to ping you when it's available.

>> For your current concern of setting up different SLs, I'd believe that
>> it could be achieved via target GUIDs as mentioned in my previous reply.
>
> Unfortunately, configuring IB QoS via target GUIDs quickly becomes too  
> complicated. As the size of clusters grow, it would require to list  
> hundreds of GUIDs in the QoS policy rules.

Yes, it's rather cumbersome at bigger scales.

Thanks,
Isaac



More information about the lustre-discuss mailing list