Re: Re: Several questions about enbd

On Monday 14 February 2005 02:47 pm, Peter T. Breuer wrote:
> Michael Ulitskiy <[email protected]> wrote:
> > I won't comment on failover script transition to mdadm now, as I'm not
> > familiar enough with mdadm itself. Nevertheless it's on the list and I'll give it a try
> 
> It's a bit more handy.  Using raid* tools means you are playing with the
> contents of a raidtab, and relying on the tols interpretation of it.
> Mdadm is a direct hands-on experience in comparision!
> 
> > as time allows (and sure I'll share the improvements if I come up with any :) ).
> > Right now I have just one question left. You said fr1 patch will read from the fastest
> > device. How the fastest device is determined? Is device speed constantly monitored
> 
> Every 1000 accesses it round-robin reads the other drives for 10
> accesses too. It maintains a rolling average of the latencies
> experienced in each access. After the short round-robin phase is up it
> shoses the drive with the lwest recorded latency for the next 1000
> accesses.
> 
> 
> > over the course of operation or it takes which device replied first during raid startup
> 
> It's retested every 1000 acceses. See above.
> 
> > and then go with it? 'Cause in latter case I can imagine it could well choose a network
> > device if it happens to reply first (think of busy systems with raid startup on the fly).
> 
> Well, I don't know if what it does is the best idea. It's just what I
> thought of for it!

I see. It's definitely better than deciding on fastest device at startup. It still seems to me 
that specifying a device priority would be a better approach though. Just a thought.

> About knowing which end was server last after both nodes come up after
> a crash ... I'm pretty convinced from thinking about it that it is
> heartbeat's business. Can you ask the heartbeat people about it? All
> the scripts get told is "start as server" or "start as client", etc.
> The decision about their role has already been made by heartbeat.

May be you're right. I'll see what I can find out as I'm still researching. On the other
hand if this is true I can't see the point of implementing DRBD's meta-data. 
From what I read so far, I doubt that heartbeat would keep its state between restarts
(when heartbeat is restarted on both nodes), at least on itself. I guess it shouldn't be
too hard to do with some additional scripting though.
Thanks for your help

> Peter
> 
> _______________________________________________
> ENBD mailing list
> [email protected]
> http://lists.community.tummy.com/mailman/listinfo/enbd
>