[Ocfs2-users] OCFS2 over DRBDv8

[Alexei_Roudnev]

Yes, o2cb will reboot one or 2 nodes.

Are you sure, that system wil recover in normal state after all this,
including ocfs2 and DRDB?

Another problem is that, in case of network glitch, you wil have both,
interconnection lost and heartbeat delay, on second node (or on all nodes).
Are you sure, that o2cb will understand such situation correctly?

I mean that using DRBD makes direct dependency between drdb device access
time and o2cb interconnections time, so
it put ocfs2 stack in untested and very unusual situation. I can bet, tht it
wil end up in cluster failure with 90% probability (if cluster lost
interconnection for relatively short time, say - 30 sconds with 12 secondr
heartbeat timeout for example).

If you read Oracle, they show that DRDB do not satisfy to requirements for
RAC cluster derundancy (in some scenarios).
OCFS2 is even wort because it will use o2cb heartbeat thru DBDB and have
much shorter timeouts (some of which are not configurable in old versions at
all).

So. It should work in the lab, allow to test many things, but don't try to
test interconnection failures on it.

----- Original Message ----- 
From: "Kilian CAVALOTTI" <kilian.cavalotti at lip6.fr>
To: "Alexei_Roudnev" <Alexei_Roudnev at exigengroup.com>
Cc: <ocfs2-users at oss.oracle.com>; <drbd-user at linbit.com>
Sent: Friday, August 18, 2006 2:28 AM
Subject: Re: [Ocfs2-users] OCFS2 over DRBDv8


On Thursday 17 August 2006 23:12, Alexei_Roudnev wrote:
> I advice against it. Even if it is possible, you definitkly will run
> into DRBD/o2cb conflict in case of any network glitch,
> and will have all your nodes self-fencing.

I'm not sure to understand the reasons of the conflict. In case of network
failure, DRBD goes split-brain, and interrupts its synchronization
procedures until manual recovery. In the mean time, o2cb will have
detected the problem and fenced both nodes. Each node will operate in
disconnected mode, that's expected behaviour, but why would DRBD and o2cb
conflict? May you please give more details?

> In addition, DRDB is not fast
> enough for OCFSv2 default settings.

May I ask why? DRBD speed is all about the interconnect link, isn't it? And
with a GbE link, which outperforms most of the not-so-new IDE/SATA
controllers, I'd appreciate any insights about how DRBD could be not fast
enough for OCFS2.

> Through, it all may be very interesting to discover in the lab.

That's actually the case, I'm not planning to base any real-world service
on such an infrastructure for now. That's more of a playground,
actually. :)

-- 
Kilian CAVALOTTI                      Administrateur réseaux et systèmes
UPMC / CNRS - LIP6 (C870)
8, rue du Capitaine Scott                          Tel. : 01 44 27 88 54
75015 Paris - France                               Fax. : 01 44 27 70 00