[Gluster-devel] Improving real world performance by moving files closer to their target workloads
Derek Price
derek at ximbiot.com
Fri May 16 14:03:47 UTC 2008
gordan at bobich.net wrote:
> Isn't that effectively the same thing? Unless there is quorum, DLM locks
> out the entire FS (it also does this when a node dies, until it gets
> definitive confirmation that it has been successfully fenced). For
> normal file I/O all nodes in the cluster have to acknowledge a lock
> before it can be granted.
Why? It requires a meta-data cache, but as long as every node in the
quorum stores a given file's most recent revision # when any lock is
granted, even if it doesn't actually sync the file data, then any quorum
should be able to agree on what the version number of the most
up-to-date copy of a file is. All nodes are required to report only if
you assume that any given file has a small number of "owners" and that
the querier doesn't know who the owner is.
To remain fault tolerant, this requires that servers make some effort to
stay up-to-date with the meta-data cache, but maybe this could be dealt
with efficiently with the DHT someone else brought up?
Regards,
Derek
--
Derek R. Price
Solutions Architect
Ximbiot, LLC <http://ximbiot.com>
Get CVS and Subversion Support from Ximbiot!
v: +1 248.835.1260
f: +1 248.246.1176
More information about the Gluster-devel
mailing list