[DRBD-user] Opteron/Xeon - Kernel 2.4/2.6 setup

Lars Ellenberg Lars.Ellenberg at linbit.com
Mon Jun 14 10:25:07 CEST 2004


/ 2004-06-14 10:12:10 +0200
\ Daniel Khan:
> Lars Ellenberg wrote:
> 
> [..]
> >>If the setup above is too risky I'll use rsync based replication for now.
> >do that.
> 
> >I think, ok, you can do it that way ...
> >but consider this:
> 
> [..]
> 
> >now, what do you think may be better wrt performance and latency?
> 
> Yes - you may be right. The only thing is that node3 is a 6 SCSI Drive /
> Raid 5 monster (Proliant DL380 G3) while node1 and node2 are smaller IDE
> machines (DL145).
> That's the only reason why I tried to put as much as possible I/O on node3.

as long as you have a mirrored device, you write on both anyways.  so
obviously your overall latency is max(latency(left disk,network,right disk)),
or worst case even sum(..(...)), and your throughput is
min(throughput(left disk,network,right disk))... 

and don't be tricked into thinking that the better the raid, the better
the sustained throughput... its more like the better the raid, the
better the reliability but often higher latency, and less sustained
(write) throughput.  (sustained meaning: you write more data for a
longer period than all those layerd caches can cope with, and so they
cannot longer hide the real underlying hardware characteristics...)

"felt performance" strongly depends on your access pattern.

	lge



More information about the drbd-user mailing list