[PVFS2-developers] PVFS2 and data redundancy

[Rob Ross]

Hi Bruce,

There is some work going on at Clemson looking into RAID-1 style 
redundancy in PVFS2, and that design is shaping up fairly nicely.  There 
is also some work going on in conjunction with NWU that is looking at 
RAID-5, but I would say that work is a little more preliminary at this 
point.

Both are relatively difficult problems.  The RAID-1 style work will rely 
on inter-server communication that is being developed now as well; that 
will help quite a bit.

There has also been some work in using dual-attached storage and commodity 
failover software to provide node failure tolerance; we've worked out many 
of the details of this here on some hardware that Dell was nice enough to 
loan us.

Both the first and the third approaches address metadata redundancy as 
well, while the second does not.  The third approach is pretty much ready 
to go now.  I would expect the first approach to mature more quickly than 
the second based on design complexity, but I wouldn't want to guess at 
when exactly it will be ready for production use.

Of course, the seemingly simplest solution would be to just add a locking
component to PVFS2, make everything work in terms of blocks, and handle
parity writes on the client side.  But then you have to make sure that the 
locking component is very reliable and you really need to journal changes 
on clients in case they fail mid-way through an operation.  And of course 
your performance drops like crazy for many common parallel write patterns 
in this case.  It would, however, be an interesting data point to have.

Hope this helps,

Rob

On Thu, 24 Jun 2004, Bruce Schwartz wrote:

> I noticed some discussion a while back (2002) about about adding
> RAID-1 and/or RAID-5 style redundancy to PVFS.  Has any design or
> further discussion occurred since then on that topic?  If not, would
> anyone hazard a guess a to how difficult that would be to add to the
> current PVFS2 codebase?