I have seen a product for Netware servers called 'Vinca StandBy Server'. It
was pretty slick. It required a network connection between servers (we used
100Mbit), and it loads a special RAID NLM into both servers. When the main
server failed, the standby would reboot and use the duplicate of the RAID
system.
It would seem this is what most people on this list are suggesting -- I just
wanted to let people know this technique has actually been used.
I would think with Linux, you could channel bond several NICs to create a
very fast direct link between primary and standby servers.
Laters,
--Bryan
-----Original Message-----
From: [EMAIL PROTECTED]
[mailto:[EMAIL PROTECTED]]On Behalf Of Jakob Østergaard
Sent: Monday, October 04, 1999 7:11 PM
To: [EMAIL PROTECTED]
Subject: Re: networked RAID-1
On Mon, Oct 04, 1999 at 04:55:04PM -0600, Brian Grossman wrote:
>
> Has anyone tried running sw-raid1 over linux's network block device? I've
> been tempted, but not quite tempted enough yet.
I think I saw that on the lkml. There were some problems (mainly
performance
under heavy load, IIRC). But it should be doable. Especially if anyone
needs
the feature ;)
> Does sw-raid1 require synchronous data xfer for the block devices
involved?
> If so, can sw-raid1 be told to not require synchronous data xfer? Does it
> make any sense to do so?
When a block device says it's done, the RAID layer can assume nothing else.
Whether the NBD code says it's done when the data reached their destination,
or when they reached some cache (or the wire), I don't know.
> Could sw-raid1 do reads only from one device, but do writes to both (for
> performance on nwb)?
That would require tweaking in the RAID code.
But if you plug in a FE card in each machine dedicated to the NBD (or one
card for each NBD) you have performance similar to that of normal disks.
Around 12 MB/s read/write (simultaneously - that's even better than most
disks - but I'm not sure that matters a lot).
> Something else I've been thinking about, but haven't had time to
> investigate properly, is intercepting filesystem calls and shipping them
> off to another machine, where they are duplicated. Anybody here have an
> idea how difficult that would be? Could it be done in a fs-independent
way?
Ouch. That could get really really nasty. RPC in the kernel is nasty
already.
Doing RAID over NBDs is pretty close to you approach, as I see it, and it's
fairly clean.
There are still implications though... You must assure that only one
machine
at a time will have the filesystem mounted, or you'll see some strange stuff
happening to your data :)
Besides, if the machine that had the filesystem mounted crashes, the
filesystem
will need a (argh, don't say it!) fsck! I guess this is the real
showstopper.
Something like PVFS (http://ece.clemson.edu/parl/pvfs/index.html) seems like
the right thing to use. Unfortunately however, it seems as though they do
not implement redundancy :(
It should be doable to write a user-space daemon that could monitor other
hosts,
and when one decides to stop responding, the NBD imported from that host
will
be marked as ``failed'' in the RAID, and our new host will fsck and mount
the
filesystem.
It would probably be wise to look into ReiserFS or ext3fs, or any other
filesystem
with journalling support. The fsck is a bad one, unless I've overlooked
something.
................................................................
: [EMAIL PROTECTED] : And I see the elder races, :
:.........................: putrid forms of man :
: Jakob Østergaard : See him rise and claim the earth, :
: OZ9ABN : his downfall is at hand. :
:.........................:............{Konkhra}...............:
