On Thu, 2009-12-10 at 16:58 +0100, Florian Haas wrote:
> Lars beat me to this but anyway...
> 
> On 2009-12-10 13:23, Beck Ingmar wrote:
> > 
> > Hi Florian,
> > 
> > why improves bonding of the replication-connection not the performance.
> 
> I did not say that. :) I said round-robin bonding _over 4 NICs_ would be
> detrimental to performance.
> 
> > We currently have a direct connection for the 1 Gb DRBD replication and
> > I
> > wanted to activate an active-active bonding (with a further LAN-port). I
> > would expect nearly double of replication-performance.
> 
> All of the following applies to DRBD Connected in protocol C:
> 
> - With 1 GBit Ethernet link, you can do about 110 MB/s (provided your
> storage subsystem has that bandwidth);
> 
> - With 2 GBit links bonded in round-robin mode, you top out at around
> 160-170 MB/s (even then you may have to tune kernel TCP parameters);
> 
> - with 3 Gbit links, your throughput will drop down to around 110 MB/s;
> 
> - with 4 links or more, you will get _significantly less_ throughput
> than a single unbonded link.
> 
> Reason for all this: TCP reordering.
> 
> > Or depends it on whether the connection between the nodes directly or
> > via 1-n switches.
> 
> I am talking about back-to-back connections. Switches could make it worse.
> 

Do Jumbo Frames make any difference here?

Mark.

-- 
Mark Watts BSc RHCE MBCS
Senior Systems Engineer, Managed Services Manpower
www.QinetiQ.com
QinetiQ - Delivering customer-focused solutions
GPG Key: http://www.linux-corner.info/mwatts.gpg

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
drbd-user mailing list
[email protected]
http://lists.linbit.com/mailman/listinfo/drbd-user

Reply via email to