[prev in list] [next in list] [prev in thread] [next in thread] 

List:       evms-devel
Subject:    [Evms-devel] Re: GFS requirements for clustered Volume Manager.
From:       David Teigland <teigland () sistina ! com>
Date:       2001-07-03 17:17:29
[Download RAW message or body]

On Fri, Jun 29, 2001 at 10:09:30AM -0700, Ramachandra Pai wrote:
> Hi Everybody,
> 	I am consolidating a list of requirements for EVMS to
> support clustered volume management.  One of 
> the areas that I am particularly interested          
> in is finding out what support clustered filesystems
> need from EVMS in order to be able to work with EVMS. 

Great, I think you'll find we're not too demanding :-)


> I have already talked to the GPFS                    
> guys here in IBM.  They told me that they would like
> to see the following things implemented in EVMS so
> that GPFS could work with EVMS:
> 
> 	1. Global volume device naming - 
>               i.e. volumes should have the same name
>               on all the nodes

Yes.


> 	2.  Persistant names - i.e. the name of a
>               volume should not change 
>               for the life of the volume, unless 
>               specifically changed by the user.

Yes.


> 	3. Large I/O path - i.e. large ( >= 256 KB)  
>            raw I/O requests actually reach the device
>            driver without being broken up into smaller
>            requests.

I'm not sure exactly what they're talking about here.  GFS won't send the
volume manager any requests larger than the page size.  There shouldn't
be any reason to break up these 4k (8k on alpha) chunks unless your 
stripe size is less than that for some odd reason.


> 	4. Scalability - i.e. support large clusters               
>            (support at least 1024 nodes where each node may have either 
>             private storage, access to shared storage, or both )

Yes.


>                        
>     	What features would GFS expect from a clustered volume manager?
>        	Clustered volume snapshots?

No, volume level snapshots will probably not be possible to use with GFS.


> 	Volume access statistics?

Not needed for GFS.  Statistics would certainly be nice, though.


> DMEP semantics on volumes?

No, GFS will not use the cluster volume manager for DMEP.  Pool will be 
used along side the cluster volume manager just for accessing DMEP on devices.

Something very basic which was not explicitly stated is that changes to
volume configuration or properties should be synchronized among machines.
i.e. handle things like two machines trying to expand a shared volume at once.

Probably obvious, but in general, be aware during the design that volumes on FC
SANs will be directly shared and possibly manipulated by several hosts.

Also notice that none of these things are really requirements of just GFS - 
all the "yes" things would be basic features for a cluster volume manager 
regardless of the file system being used.

-- 
Dave Teigland  <teigland@sistina.com>

[prev in list] [next in list] [prev in thread] [next in thread]