[Beowulf] Sun buys Lustre

[Mark Hahn]

> I'm not sure if everyone has seen this yet, but Sun just bought Cluster
> File System (Lustre):

wow, that's a bit of a shock.  I had heard mutterings about CFS using ZFS,
but would not have guessed a buyout.  where does that leave HP, for
instance?  or even SGI, who seem to now be promoting Lustre for their 
new blade clusters...

the combination of ZFS and Luster is a bit strange to contemplate:

 	- Lustre's main thing is to set up client-OST parallel streams
 	and then get out of the way.  it uses a single metadata server,
 	which can be a bottleneck for non-giant-stream workloads, and
 	which currently stores per-file ost-mapping info via EAs in an
 	ext3 FS.  OST'S don't need much functionality in their local FS,
 	and use ext3 (with extents) as well.

 	- ZFS's main thing is managing storage pools directly so as to
 	provide intelligent file-level redundancy.  and it has comfortingly
 	high size limits and some nice integrity features.  but afaik,
 	no provisions at all for clustering or parallel access.

I guess you could broadly compare CFS's OST's to ZFS's storage pools
(sorry, don't know the ZFS terminology).  but "clusterizing" the metadata
logic of ZFS (parallelizing it and presumably distributing it across clients)
sounds kind of tricky.

(I confess my only experience is with HP's SFS version of Luster, and none 
at all with ZFS.)