[Linux-HA] High availability, fast, fileservers

[Lars Ellenberg]

/ 2004-03-17 22:32:00 -0500
\ Jeff Tucker:
> >>Finally, a question about DRDB. Is it reliable enough that you'd be
> >>willing  to set up a DRDB pair with each machine running a single RAID-0
> >>array (and  maybe the OS is on a different disk)? In this case, a single
> >>drive failure  would fail the entire array and the DRDB would need to
> >>fail over to the  other machine. This could happen as often as once or
> >>twice a year if disk  failures are about what I expect.
> >>
> >
> >In my limited drbd experience (a nice little redundant samba server), it
> >was most advantageous to split the storage in two, and have each server
> >host half of it. When the failover happened, things slowed down on the
> >poor box that was left to serve both, but kept working well enough to
> >allow repair of the dead box without panic.
> >
> 
> In reading over the archives of this list, there was a comment posted, 
> maybe months ago, that an active-active setup like that might have 
> occasional weird problems. One of the DRDB developers indicated he thought 
> they might still have some races and that this setup would be more likely 
> to hit them.

That was most likely me...
I *think* with DRBD 0.6.12 we got rid of that one thing that could
lead to a distributed deadlock when one had high memory pressure
and high IO load at the same time on both nodes...

So I'd even recommend such a setup, if the data split comes natural.

> Obviously, this lets you spread reads across both boxes and should nearly 
> double your read performance. For writes, data still needs to be written to 
> both systems eventually, so I'd think writes aren't appreciably helped by 
> making each system active for half the data. But, spreading out the reads 
> might help a lot. Do others run in this configuration?

IIRC, some (most?) of the servers managed directly by LinBit for
our customers are configured this way. And there we did never have
these "weird" problems, even when they still where theoretically possible...
I don't know the exact details, though... only loosely affiliated
:)

	Lars Ellenberg