I'm a little confused by your explanation of 'just do the bonding at the
guest level'. I apologize for my ignorance here, but I'm trying to prepare
myself for a similar configuration where I'm going to need to get all much
bandwidth out of the bond as possible. How would bonding multiple
interfaces at the VM level provide a better balance than at the hypervisor
level? Wouldn't the traffic more or less end up traveling the same path
regardless of the virtual interface?
I'm trying to plan out an oVirt implementation where I would like to bond
multiple interfaces on my hypervisor nodes for balancing/redundancy, and
I'm very curious what others have done with Cisco hardware (in my case, a
pair of 3650's with MEC) in order to get the best solution.
I will read through these threads and see if I can gain a better
understanding, but if you happen to have an easy explanation that would
help my understand, I would greatly appreciate it.
On Wed, Dec 31, 2014 at 1:01 AM, Blaster wrote:
>
> Thanks for your thoughts. The problem is, most of the data is transmitted
> from a couple apps to a couple systems. The chance of a hash collision
> (i.e., most of the data going out the same interface anyway) is quite
> high. On Solaris, I just created two physical interfaces each with their
> own IP, and bound the apps to the appropriate interfaces. This worked
> great. Imagine my surprise when I discovered this doesn’t work on Linux
> and my crash course on weak host models.
>
> Interesting that no one commented on my thought to just do the bonding at
> the guest level (and use balance-alb) instead of at the hypervisor level.
> Some ESXi experts I have talked to say this is actually the preferred
> method with ESXi and not to do it at the hypervisor level, as the VM knows
> better than VMware.
>
> Or is the bonding mode issue with balance-alb/tlb more with the Linux TCP
> stack itself and not with oVirt and VDSM?
>
>
>
> On Dec 30, 2014, at 4:34 AM, Nikolai Sednev wrote:
>
> Mode 2 will do the job the best way for you in case of static LAG
> supported only at the switch's side, I'd advise using of xmit_hash_policy
> layer3+4, so you'll get better distribution for your DC.
>
>
> Thanks in advance.
>
> Best regards,
> Nikolai
>
> Nikolai Sednev
> Senior Quality Engineer at Compute team
> Red Hat Israel
> 34 Jerusalem Road,
> Ra'anana, Israel 43501
>
> Tel: +972 9 7692043
> Mobile: +972 52 7342734
> Email: nsed...@redhat.com
> IRC: nsednev
>
> --
> *From: *users-requ...@ovirt.org
> *To: *users@ovirt.org
> *Sent: *Tuesday, December 30, 2014 2:12:58 AM
> *Subject: *Users Digest, Vol 39, Issue 173
>
> Send Users mailing list submissions to
> users@ovirt.org
>
> To subscribe or unsubscribe via the World Wide Web, visit
> http://lists.ovirt.org/mailman/listinfo/users
> or, via email, send a message with subject or body 'help' to
> users-requ...@ovirt.org
>
> You can reach the person managing the list at
> users-ow...@ovirt.org
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of Users digest..."
>
>
> Today's Topics:
>
>1. Re: ??: bond mode balance-alb (Jorick Astrego)
>2. Re: ??: bond mode balance-alb (Jorick Astrego)
>3. HostedEngine Deployment Woes (Mikola Rose)
>
>
> --
>
> Message: 1
> Date: Mon, 29 Dec 2014 20:13:40 +0100
> From: Jorick Astrego
> To: users@ovirt.org
> Subject: Re: [ovirt-users] ??: bond mode balance-alb
> Message-ID: <54a1a7e4.90...@netbulae.eu>
> Content-Type: text/plain; charset="utf-8"
>
>
> On 12/29/2014 12:56 AM, Dan Kenigsberg wrote:
> > On Fri, Dec 26, 2014 at 12:39:45PM -0600, Blaster wrote:
> >> On 12/23/2014 2:55 AM, Dan Kenigsberg wrote:
> >>> Bug 1094842 - Bonding modes 0, 5 and 6 should be avoided for VM
> networks
> >>> https://bugzilla.redhat.com/show_bug.cgi?id=1094842#c0
> >> Dan,
> >>
> >> What is bad about these modes that oVirt can't use them?
> > I can only quote jpirko's workds from the link above:
> >
> > Do not use tlb or alb in bridge, never! It does not work, that's it.
> The reason
> > is it mangles source macs in xmit frames and arps. When it is
> possible, just
> > use mode 4 (lacp). That should be always possible because all
> enterprise
> > switches support that. Generally, for 99% of use cases, you *should*
> use mode
> > 4. There is no reason to use other modes.
> >
> This switch is more of an office switch and only supports part of the
> 802.3ad standard:
>
>
> PowerConnect* *2824
>
> Scalable from small workgroups to dense access solutions, the 2824
> offers 24-port flexibility plus two combo small?form?factor
> pluggable (SFP) ports for connecting the switch to other networking
> equipment located beyond the 100 m distance limitations of copper
> cabling.
>
> Industry-standard link aggregation ad