[Pvfs2-developers] bmi multiple address endpoints

[Sam Lang]

Hi Pete,

I was actually thinking the tabfile list would be optional, and its  
use would be only to filter out addresses in the list provided for  
each Alias.  The BMI methods would still be enabled based on what was  
specified in the BMIModules config option.  This still hurts  
performance because of the sequential polling with multiple methods  
enabled, but I was thinking that could be solved separately (multiple  
threads, etc.).  Should we try to not enable methods specified in  
BMIModules if they're filtered out with the mntent option?

-sam


On Nov 9, 2007, at 9:13 AM, Pete Wyckoff wrote:

> slang at mcs.anl.gov wrote on Wed, 07 Nov 2007 16:21 -0600:
>> I discussed the desired behavior we want out of this fail-over  
>> code with
>> folks offline, and we came up with a plan.
>
> Sorry for the delay.  Head down in SC.  I definitely like this plan
> and believe you addressed the common cases on server and client.
> Murali's ideas of protocol splitting are cool, but I think would be
> too complex for now, when we just want to get something working for
> failover.
>
> The client mount option.  We already have a fs.conf BMIModules that
> affects server and all potential clients.  You want to add a
> client-specific list to its pvfs2tab that limits that.  This may be
> something that also should be more general.  We might get into more
> messes where different clients call hosts different things, and
> client 27 shouldn't even try to connect to hostb as it will go over
> the wrong network, or such.
>
> But, for now, a tab-file list of bmi addresses is expedient, maybe
> with an eye for a generic "fs.conf client-specific override"
> mechanism in the future?  I caught on this because it will be
> "BMIModules" in fs.conf, but "bmi" in pvfs2tab.  Would like those to
> be the same.
>
> 		-- Pete
>