On Sep 13, 2007, at 01:08, Brandorr wrote: > Lustre is here today, and very widely deployed. No need to wait for > pNFS. (Plus someone else would have snapped them up as well). > > I see the biggest win for some is getting the dev team, and HPC > customers. QFS is not HPC, it's more archival... (Or am I thinking > SAMFS)
err - no .. QFS is very HPC oriented and easily combined with SAM (the Storage Archive Management component) for offloading content in an easily retrievable archive - a big difference here is Lustre will be plugging into the ZFS framework at the DMU and provides transactional dispersion across a grid extending the ZFS object framework. QFS is well suited for shared storage that's direct attached with many more tunable options to turn off or fine-tune the caching and block layout instead of adopting the "everything is transactional" approach. If you're already optimized on your stream, there's no sense in passing through unnecessary layers to do additional buffering and caching. If you're not and want the extra functionality that can be done on the transactional layer as data passes through a CoW then ZFS makes sense. To see them all working together happily, take a look at the canned solution: http://www.sun.com/servers/cr/scalablestorage/specs.xml You might also want to take a look at the T10 OSD interface project (opensolaris.org/os/project/osd) which is a project aimed more at providing a framework for the storage level object interface using the T10 object definitions (ie: http://www.t10.org/ftp/t10/drafts/ osd2/osd2r02.pdf) which is slightly different than ZFS objects. --- .je
