Software RAID issues (was Re: Suggestions solicited, server bring up)

[Bill McGonigle]

> I don't want to go commercial, so I won't guess the name, but are the
> initials BFCC, but chance? ;-)

you'll have to check when the new website gets pushed to live. :)

> A fellow on this list at the Birthday party said that iSCSI had a lot less
> network overhead and much better real throughput than NFS.  Is there a way
> to bring NFS closer to even?  Parallelization does not quite address the
> issue of overhead.  Is there something else in pNFS that does?

iSCSI and NFS seem to be pretty similar in raw throughput, but iSCSI does
much better when it comes to frequent metadata updates due to NFS
limitations.  AIUI, pNFS fixes this problem, among others, by allowing
more to happen in parallel.  There could be some workloads where the files
are much smaller than blocksize in which NFS could come out ahead.

But the real benefits of doing NFS are for multiple-access scenarios. 
e.g. on the host you can see the files, so you can back them up natively. 
Or you can have multiple systems coming in r/w on the same NFS export. 
Not necessary or feasible for every scenario, but real nice when you can
tolerate NFS.

> My goal
> is
> to get our database storage into our private cloud storage (under
> development), but we get >1.5GBps now on some of our FusionIO stores, so
> even with 10Gbps NICs, we would be taking a step down before accouting for
> overhead.

yeah, NFS and databases aren't really a great mapping - not enough
semantics are supported even if they were fast enough.  Does your database
support multi-level storage (e.g. putting your WAL or cache on your
fastest drives)?  FYI, ZFS supports this on the back-end (fast cache
drives in front of cheaper slower drives.  But odds are NAS/SAN is slower
than RAM on a local bus. :)

> What?! No. Oracle GPL'ing?  Really?  Got any articles handy were I can
> read
> up on the details. That's very exciting.

  http://oss.oracle.com/projects/

btrfs is already GPL, so I'm having trouble figuring out why they would
insist on keeping ZFS CDDL.

-Bill