Doing bonding without LACP is probably going to end up being painful.
Sooner or later you're going to end up with one end thinking that
bonding is working while the other end thinks that it's not, and half
of your traffic is going to get black-holed.
I've had moderately decent luck running Ceph on top of a weird
network by carefully controlling the source address that every
outbound connection uses and then telling Ceph that it's running with
a 1-network config. With Linux, the default source address of an
outbound TCP connection is a function of the route that the kernel
picks to send traffic to the remote end, and you can override it on a
per-route basis (it's visible as the the 'src' attribute in iproute).
I have a mixed Infiniband+GigE network with each host running an
OSPF routing daemon (for non-Ceph reasons, mostly), and the only two
ways that I could get Ceph to be happy were:
1. Turn off the Infiniband network. Slow, and causes other problems.
2. Tell Ceph that there was no cluster network, and tell the OSPF
daemon to always set src=$eth0_ip on routes that it adds. Then just
pretend that the Ethernet network is the only one that exists, and
sometimes you get a sudden and unexpected boost in bandwidth due to
/32 routes that send traffic via Infiniband instead of Ethernet.
It works, but I wouldn't recommend it for production. It would have
been cheaper for me to buy a 10 GigE switch and cards for my garage
than to have debugged all of this, and that's just for a hobby project.
OTOH, it's probably the only way to get working multipathing for Ceph.
On Thu, Jun 5, 2014 at 10:50 AM, Cedric Lemarchand
<[email protected] <mailto:[email protected]>> wrote:
Le 05/06/2014 18:27, Sven Budde a écrit :
> Hi Alexandre,
>
> thanks for the reply. As said, my switches are not stackable,
so using LCAP seems not to be my best option.
>
> I'm seeking for an explanation how Ceph is utilizing two (or
more) independent links on both the public and the cluster network.
AFAIK, Ceph do not support multiple IP link in the same "designated
network" (aka client/osd networks). Ceph is not aware of links
aggregations, it has to be done at the Ethernet layer, so :
- if your switchs are stackable, you can use traditional LACP on both
sides (switch and Ceph)
- if they are not, and as Mariusz said, use the appropriate
bonding mode
on the Ceph side and do not use LCAP on switchs.
More infos here :
http://www.linuxfoundation.org/collaborate/workgroups/networking/bonding
Cheers !
>
> If I configure two IPs for the public network on two NICs, will
Ceph route traffic from its (multiple) OSDs on this node over
both IPs?
>
> Cheers,
> Sven
>
> -----Ursprüngliche Nachricht-----
> Von: Alexandre DERUMIER [mailto:[email protected]
<mailto:[email protected]>]
> Gesendet: Donnerstag, 5. Juni 2014 18:14
> An: Sven Budde
> Cc: [email protected] <mailto:[email protected]>
> Betreff: Re: [ceph-users] Ceph networks, to bond or not to bond?
>
> Hi,
>
>>> My low-budget setup consists of two gigabit switches, capable
of LACP,
>>> but not stackable. For redundancy, I'd like to have my links
spread
>>> evenly over both switches.
> If you want to do lacp with both switches, they need to be
stackable.
>
> (or use active-backup bonding)
>
>>> My question where I didn't find a conclusive answer in the
>>> documentation and mailing archives:
>>> Will the OSDs utilize both 'single' interfaces per network, if I
>>> assign two IPs per public and per cluster network? Or will
all OSDs
>>> just bind on one IP and use only a single link?
> you just need 1 ip by bond.
>
> with lacp, the load balacing use an hash algorithm, to
loadbalance tcp connections.
> (that also mean than 1 connection can't use more than 1 link)
>
> check that your switch support ip+port hash algorithm,
(xmit_hash_policy=layer3+4 is linux lacp bonding)
>
> like this, each osd->osd can be loadbalanced, same for your
clients->osd.
>
>
>
>
>
>
> ----- Mail original -----
>
> De: "Sven Budde" <[email protected]
<mailto:[email protected]>>
> À: [email protected] <mailto:[email protected]>
> Envoyé: Jeudi 5 Juin 2014 16:20:04
> Objet: [ceph-users] Ceph networks, to bond or not to bond?
>
> Hello all,
>
> I'm currently building a new small cluster with three nodes,
each node having 4x 1 Gbit/s network interfaces available and
8-10 OSDs running per node.
>
> I thought I assign 2x 1 Gb/s for the public network, and the
other 2x 1 Gb/s for the cluster network.
>
> My low-budget setup consists of two gigabit switches, capable
of LACP, but not stackable. For redundancy, I'd like to have my
links spread evenly over both switches.
>
> My question where I didn't find a conclusive answer in the
documentation and mailing archives:
> Will the OSDs utilize both 'single' interfaces per network, if
I assign two IPs per public and per cluster network? Or will all
OSDs just bind on one IP and use only a single link?
>
> I'd rather avoid bonding the NICs, as if one switch fails,
there would be at least one node unavailable, in worst case 2
(out of 3) ...rendering the cluster inoperable.
>
> Are there other options I missed? 10 GE is currently out of our
budget ;)
>
> Thanks,
> Sven
>
>
> _______________________________________________
> ceph-users mailing list
> [email protected] <mailto:[email protected]>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>
> _______________________________________________
> ceph-users mailing list
> [email protected] <mailto:[email protected]>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
--
Cédric
_______________________________________________
ceph-users mailing list
[email protected] <mailto:[email protected]>
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
