[ILUG] NBD's, lihux-HA, GFS/GNBD vs. DRBD/heartbeat

[Paul Jakma]

On Mon, 28 Feb 2005, John Moylan wrote:

> It is nice also that NBD is a mature part of the kernel (first
> incorporated back in 2.1), from my research DRBD is really only a bunch

> Tyan Opteron SMP Mobo with PCI-X
> 2 Opteron 240's

Sounds to me like you intend these systems to IO bound. Tip of the 
day:

- dont use SMP for IO bound tasks, unless you know it would be a
   benefit. A second CPU waiting for IO usually makes little
   difference..

- avoiding SMP automatically removes several classes of, SMP
   specific, reliability-affecting, bugs.
   (this applies particularly when you're going for a more unusual
    setup).

In essence:

- dont spec SMP just cause you can.

> striped in hardware and then mirrored onto the second card using S/W RAID.
> 2* Gb Ethernet (channel bonded)

Go for routing at layer-3 instead, if at all possible. You're adding 
yet another stack of not-widely-used kernel code into the mix by 
going the channel bonding route. Not-widely-used kernel code is often 
antithetical to reliability[1].

(user mode code often is too, indeed software in general seems to be, 
but at least user mode code can be restarted without too much fuss, 
possibly even by a simple watchdog script.).

1. Oh, let "who wrote the driver?" be your guide when selecting 
hardware, if reliability is a concern, along with "And did the vendor 
provide specs?". Crap, cheap hardware is sometimes a safer bet if the 
drivers were written by a kernel guru of long-standing (particularly 
if he/she had access to specs) than hot-poop hardware written by 
vendor staffers..

PS: Apparently ENBD:

" ..... is an industrial-strength version of the Linux kernel NBD. "

see: http://www.it.uc3m.es/ptb/enbd/

no idea whether there's truth in that claim.

> John

regards,
-- 
Paul Jakma	paul at clubi.ie	paul at jakma.org	Key ID: 64A2FF6A
Fortune:
Children are natural mimics who act like their parents despite every
effort to teach them good manners.