Re: Ceph with Clos IP fabric

[Aaron Bassett <Aaron.Bassett@xxxxxxxxxxxxx>]

Agreed. In an ideal world I would have interleaved all my compute, long term storage and processing posix. Unfortunately, business doesn't always work out so nicely so I'm left with buying and building out to match changing needs. In this case we are a small part of a larger org and have been allocated X racks in the cage, which is at this point land locked with no room to expand so it is actual floor space that's limited. Hence the necessity to go as dense as possible when adding any new capacity. Luckily ceph is flexible enough to function fine when deployed like an EMC solution, it's just muuuch cheaper and more fun to operate!

Aaron 

On Apr 24, 2017, at 12:59 AM, Richard Hesse <richard.hesse@xxxxxxxxxx> wrote:

It's not a requirement to build out homogeneous racks of ceph gear. Most larger places don't do that (it creates weird hot spots).  If you have 5 racks of gear, you're better off spreading out servers in those 5 than just a pair of racks that are really built up. In Aaron's case, he can easily do that since he's not using a cluster network.

Just be sure to dial in your crush map and failure domains with only a pair of installed cabinets.

Thanks for sharing Christian! It's always good to hear about how others are using and deploying Ceph, while coming to similar and different conclusions.

Also,when you say datacenter space is expensive, are you referring to power or actual floor space? Datacenter space is almost always sold by power and floor space is usually secondary. Are there markets where that's opposite? If so, those are ripe for new entrants! 

On Apr 23, 2017 7:56 PM, "Christian Balzer" <chibi@xxxxxxx> wrote:

Hello,

Aaron pretty much stated most of what I was going to write, but to
generalize things and make some points more obvious, I shall pipe up as
well.

On Sat, 22 Apr 2017 21:45:58 -0700 Richard Hesse wrote:

> Out of curiosity, why are you taking a scale-up approach to building your
> ceph clusters instead of a scale-out approach? Ceph has traditionally been
> geared towards a scale-out, simple shared nothing mindset.

While true, scale-out does come at a cost:
a) rack space, which is mighty expensive where we want/need to be and also
of limited availability in those locations.
b) increased costs by having more individual servers, as in having two
servers with 6 OSDs versus 1 with 12 OSDs will cost you about 30-40% more
at the least (chassis, MB, PSU, NIC).

And then there is the whole scale thing in general, I'm getting the
impression that the majority of Ceph users have small to at best medium
sized clusters, simply because they don't need all that much capacity (in
terms of storage space).

Case in point, our main production Ceph clusters fit into 8-10U with 3-4
HDD based OSD servers and 2-4 SSD based cache tiers, obviously at this
size with everything being redundant (switches, PDU, PSU).
Serving hundreds (nearly 600 atm) of VMs, with a planned peak around
800 VMs.
That Ceph cluster will never have to grow beyond this size.
For me Ceph (RBD) was/is a more scalable approach than DRBD, allowing for
n+1 compute node deployments instead of having pairs (where one can't live
migrate to outside of this pair).

>These dual ToR
> deploys remind me of something from EMC, not ceph. Really curious as I'd
> rather have 5-6 racks of single ToR switches as opposed to three racks of
> dual ToR. Is there a specific application or requirement? It's definitely
> adding a lot of complexity; just wondering what the payoff is.
>

If you have plenty of racks, bully for you.
Though personally I'd try to keep failure domains (especially when they
are as large as full rack!) to something like 10% of the cluster.
We're not using Ethernet for the Ceph network (IPoIB), but if we were it
would be dual TORS with MC-LAG (and dual PSU, PDU) all the way.
Why have a SPOF that WILL impact your system (a rack worth of data
movement) in the first place?

Regards,

Christian

> Also, why are you putting your "cluster network" on the same physical
> interfaces but on separate VLANs? Traffic shaping/policing? What's your
> link speed there on the hosts? (25/40gbps?)
>
> On Sat, Apr 22, 2017 at 12:13 PM, Aaron Bassett <Aaron.Bassett@xxxxxxxxxxxxx
> > wrote:
>
> > FWIW, I use a CLOS fabric with layer 3 right down to the hosts and
> > multiple ToRs to enable HA/ECMP to each node. I'm using Cumulus Linux's
> > "redistribute neighbor" feature, which advertises a /32 for any ARP'ed
> > neighbor. I set up the hosts with an IP on each physical interface and on
> > an aliased looopback: lo:0. I handle the separate cluster network by adding
> > a vlan to each interface and routing those separately on the ToRs with acls
> > to keep traffic apart.
> >
> > Their documentation may help clarify a bit:
> > https://docs.cumulusnetworks.com/display/DOCS/Redistribute+
> > Neighbor#RedistributeNeighbor-ConfiguringtheHost(s)
> >
> > Honestly the trickiest part is getting the routing on the hosts right, you
> > essentially set static routes over each link and the kernel takes care of
> > the ECMP.
> >
> > I understand this is a bit different from your setup, but Ceph has no
> > trouble at all with the IPs on multiple interfaces.
> >
> > Aaron
> >
> > Date: Sat, 22 Apr 2017 17:37:01 +0000
> > From: Maxime Guyot <Maxime.Guyot@xxxxxxxxx>
> > To: Richard Hesse <richard.hesse@xxxxxxxxxx>, Jan Marquardt
> > <jm@xxxxxxxxxxx>
> > Cc: "ceph-users@xxxxxxxxxxxxxx" <ceph-users@xxxxxxxxxxxxxx>
> > Subject: Re: Ceph with Clos IP fabric
> > Message-ID: <919C8615-C50B-4611-9B6B-13B4FBF69C61@xxxxxxxxx>
> > Content-Type: text/plain; charset="utf-8"
> >
> > Hi,
> >
> > That only makes sense if you're running multiple ToR switches per rack for
> > the public leaf network. Multiple public ToR switches per rack is not very
> > common; most Clos crossbar networks run a single ToR switch. Several
> > >guides on the topic (including Arista & Cisco) suggest that you use
> > something like MLAG in a layer 2 domain between the switches if you need
> > some sort of switch redundancy inside the rack. This increases complexity,
> > and most people decide that it's not worth it and instead scale out across
> > racks to gain the redundancy and survivability that multiple ToR offer.
> >
> > If you use MLAG for L2 redundancy, you?ll still want 2 BGP sessions for L3
> > redundancy, so why not skipping the MLAG all together and terminating your
> > BGP session on each ToR?
> >
> > Judging by the routes (169.254.0.1), you are using BGP unnumebered?
> >
> > It sounds like the ?ip route get? output you get when using dummy0 is
> > caused by a fallback on the default route, supposedly on eth0? Can check
> > the exact routes received on server1 with ?show ip bgp neighbors <neighbor>
> > received-routes? once you enable ?neighbor <neighbor> soft-reconfiguration
> > inbound? and what?s installed in the table ?ip route??
> >
> >
> > Intrigued by this problem, I tried to reproduce it in a lab with
> > virtualbox. I ran into the same problem.
> >
> > Side note: Configuring the loopback IP on the physical interfaces is
> > workable if you set it on **all** parallel links. Example with server1:
> >
> > ?iface enp3s0f0 inet static
> >  address 10.10.100.21/32
> > iface enp3s0f1 inet static
> >  address 10.10.100.21/32
> > iface enp4s0f0 inet static
> >  address 10.10.100.21/32
> > iface enp4s0f1 inet static
> >  address 10.10.100.21/32?
> >
> > This should guarantee that the loopback ip is advertised if one of the 4
> > links to switch1 and switch2 is up, but I am not sure if that?s workable
> > for ceph?s listening address.
> >
> >
> > Cheers,
> > Maxime
> >
> > From: Richard Hesse <richard.hesse@xxxxxxxxxx>
> > Date: Thursday 20 April 2017 16:36
> > To: Maxime Guyot <Maxime.Guyot@xxxxxxxxx>
> > Cc: Jan Marquardt <jm@xxxxxxxxxxx>, "ceph-users@xxxxxxxxxxxxxx" <
> > ceph-users@xxxxxxxxxxxxxx>
> > Subject: Re: Ceph with Clos IP fabric
> >
> > On Thu, Apr 20, 2017 at 2:13 AM, Maxime Guyot <Maxime.Guyot@xxxxxxxxx<

> > mailto:Maxime.Guyot@xxxxxxxxx <Maxime.Guyot@xxxxxxxxx>>> wrote:
> >
> > 2) Why did you choose to run the ceph nodes on loopback interfaces as
> > opposed to the /24 for the "public" interface?
> >
> > I can?t speak for this example, but in a clos fabric you generally want to
> > assign the routed IPs on loopback rather than physical interfaces. This way
> > if one of the link goes down (t.ex the public interface), the routed IP is
> > still advertised on the other link(s).
> >
> > That only makes sense if you're running multiple ToR switches per rack for
> > the public leaf network. Multiple public ToR switches per rack is not very
> > common; most Clos crossbar networks run a single ToR switch. Several guides
> > on the topic (including Arista & Cisco) suggest that you use something like
> > MLAG in a layer 2 domain between the switches if you need some sort of
> > switch redundancy inside the rack. This increases complexity, and most
> > people decide that it's not worth it and instead  scale out across racks to
> > gain the redundancy and survivability that multiple ToR offer.
> >
> > On Thu, Apr 20, 2017 at 4:04 AM, Jan Marquardt <jm@xxxxxxxxxxx<mailto:jm@

> > artfiles.de <jm@xxxxxxxxxxx>>> wrote:
> >
> > Maxime, thank you for clarifying this. Each server is configured like this:
> >
> > lo/dummy0: Loopback interface; Holds the ip address used with Ceph,
> > which is announced by BGP into the fabric.
> >
> > enp5s0: Management Interface, which is used only for managing the box.
> > There should not be any Ceph traffic on this one.
> >
> > enp3s0f0: connected to sw01 and used for BGP
> > enp3s0f1: connected to sw02 and used for BGP
> > enp4s0f0: connected to sw01 and used for BGP
> > enp4s0f1: connected to sw02 and used for BGP
> >
> > These four interfaces are supposed to transport the Ceph traffic.
> >
> > See above. Why are you running multiple public ToR switches in this rack?
> > I'd suggest switching them to a single layer 2 domain and participate in
> > the Clos fabric as a single unit, or scale out across racks (preferred).
> > Why bother with multiple switches in a rack when you can just use multiple
> > racks? That's the beauty of Clos: just add more spines if you need more
> > leaf to leaf bandwidth.
> >
> > How many OSD, servers, and racks are planned for this deployment?
> >
> > -richard
> >
> >
> > CONFIDENTIALITY NOTICE
> > This e-mail message and any attachments are only for the use of the
> > intended recipient and may contain information that is privileged,
> > confidential or exempt from disclosure under applicable law. If you are not
> > the intended recipient, any disclosure, distribution or other use of this
> > e-mail message or attachments is prohibited. If you have received this
> > e-mail message in error, please delete and notify the sender immediately.
> > Thank you.
> >
> > _______________________________________________
> > ceph-users mailing list
> > ceph-users@xxxxxxxxxxxxxx
> > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
> >
> >

--
Christian Balzer        Network/Systems Engineer
chibi@xxxxxxx           Global OnLine Japan/Rakuten Communications
http://www.gol.com/

CONFIDENTIALITY NOTICE
This e-mail message and any attachments are only for the use of the intended recipient and may contain information that is privileged, confidential or exempt from disclosure under applicable law. If you are not the intended recipient, any disclosure, distribution or other use of this e-mail message or attachments is prohibited. If you have received this e-mail message in error, please delete and notify the sender immediately. Thank you.

_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com