[DRBD-user] Procedure to convert internal to external meta data

Caspar Smit c.smit at truebit.nl
Fri Oct 15 09:08:09 CEST 2010

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.

2010/10/14 Theo Band <theo.band at greenpeak.com>:
>  On 10/13/2010 10:21 AM, Caspar Smit wrote:
>> Hi all,
>> Is there a procedure to convert internal meta data to external meta
>> data without loosing any data?
>> I was thinking of something like this:
>> 1. drbdadm down r0
>> 2. drbdadm dump-md r0 >/tmp/metadata
>> 3. edit config for r0 changing internal to external metadata.
>> 4. drbdadm create-md r0
>> 5. drbdmeta_cmd=$(drbdadm -d dump-md test-disk)
>>     ${drbdmeta_cmd/dump-md/restore-md} /tmp/metadata
>> 6. repeat steps 1-5 on the other node
>> 7. drbdadm up r0 (both nodes)
>> Is this the way to go?
> In your proposal you bring r0 down on one node, so this is not a dual
> primary setup I guess?

No it's primary/secondary indeed

 If one node is down, why not simply change your
> configuration (with a separate meta data disk) initialize and reconnect
> again. It will start resyncing which will take some time.

It's a 32 TB setup (4x8TB) so I want to prevent having to sync 32TB again :)

 But apart from
> that you will not loose any data. After the disks are in sync on both
> nodes you could do the same on the other node.
> In a dual primary setup a disconnect on one node should also work (one
> node will be diskless during this maintenance time).
> But why not stick to internal meta data?

It's 4 x (8TB software RAID5), and while the sync speed is 100MB/s per 8TB
so 400MB/s in total which is certainly not bad I was kinda disappointed in
the random IOPS i could generate on the sets, i read somewhere on a forum
that for a RAID5 MD was not a really good idea to have internal meta data
due to the many seeks needed, so i figured i could just extract the
metadata to a seprate device to gain some (random io) performance.

This should/could be a fairly simple procedure not?

Kind regards,

> Theo

More information about the drbd-user mailing list