Network mirroring (was:R: Questions)

Marcelo Tosatti marcelo at conectiva.com.br
Thu Nov 4 20:41:03 MST 1999


On Fri, 5 Nov 1999, Giampaolo Tomassoni wrote:

> > > A better fs-independent networked raid solution should be scalable, 
> allowing
> > > more than two hosts to cooperate in the creation of a 'virtual' raid. This
> > GNBD allows more than one client.
> 
> It does. But needs to open a tcp connection between two nodes. I.e.: if you 
> have 3 nodes among which you want create a raid1, you have to create a tcp conn 
> between node 1 and 2, as well as between node 2 and 3.
> 
> Besides, this way if node 2 crashes, you miss a route to sync node 1 and 3 
> toghether...
> 
> 
> > But it suffers from consistency problems (at least this is what i figured
> > out in my 1 min look at the code).
> 
> You mean at startup? Reading the sources, I had the feeling that a read and a 
> write wouldn't be completed untill the mbd completed the tcp transaction with 
> the remote node.
Remember stale data problem:
Your buffer cache may have data readed a _long_ time ago and that is not
the same now. But, as Mike Tilstra said, this is handled by GFS. 
And in your device-level case, GNBD is not the place to do this
consistency.

> > > would be better implemented using (possibly multicast) udp messages. Here 
> the
> > > matter is complicated by the need to keep in sync every local image of the 
> raid
> > > (srv A writes and sends an update, srv B looses the update and wanna read), 
> but
> > > It should be made at the md or raidX layer in Linux, not below...
> > Nod. I will send the RAID1 write-only patch tomorrow which at least makes
> > the NBD+RAID1 more faster.
> 
> Nod? What does it mean 'nod'?
I mean 'agreed'. 

  - Marcelo





More information about the Linux-HA mailing list