distributed storage

Robert Ross rross at mcs.anl.gov
Sat Jun 1 20:31:07 PDT 2002


What do you want to do with all that space?  Store large datasets?  Store
usenet news?  The expected workload has a HUGE impact on your choices,
because it says things about a) how many nodes files need to be spread
across to get the bandwidth you need, and b) what the access patterns to
the file system might look like.

You've mentioned that you are concerned about availability in the presence
of node failures.  That makes good sense.  And for PVFS, if a node that
has your file's data goes down, then you can't get to that file.

That's a little different than your description, and the difference could
be important.  If, for example, you really just want to use the space and
it's ok to only get a single link's worth of bandwidth, then you could set
up a PVFS volume that puts files on single servers.  Then any given file
has at most two nodes on which it is dependent (metadata server and the
node on which the data is stored).  Most data would still be available in
most failure cases.

Likewise you could set up the system to stripe to only two I/O servers and
so on.  It's not the ideal solution by any means, but it is a good
tradeoff between availability and performance given the packages
available.  There aren't too many choices at the "free" price point.  
Cross-mounted NFS volumes would be the other one that I can think of :).

Rob Ross, Mathematics and Computer Science Division, Argonne National Lab

On Thu, 30 May 2002, Yudong Tian wrote:

> Hi everyone!
>     I'd like to hear your expert opinion on how to effectively use the
> distributed storage capacity of a cluster. Suppose you have hundreds
> of nodes, each with a local hard disk, thus the accumulated capacity
> is huge. Aslo you have large amount of data to be stored. How can we
> reliably take advantage of the storage on the nodes?
>    I checked out PVFS. It is nice, but it seems to me the reliability
> is an issue -- if one of the storage nodes goes down, the whole FS is
> not available (correct me if I am wrong).  CXFS and GFS are not our
> options because they cost $$.  Any suggestions on how we can best use
> the capacity, while having no problems when one or two nodes crash?
> Thanks.
> -------------------------------------------------------
> Falun Dafa: Truthfulness Benevolence Forbearance
> http://www.falundafa.org 
> -------------------------------------------------------
> Yudong Tian, Ph.D. GSFC/NASA  (301) 504-5825

More information about the Beowulf mailing list