[CentOS] bonding theory question

Sat Sep 6 17:11:41 UTC 2008
Mag Gam <magawake at gmail.com>


Thankyou! Your explanation helps a lot. Its makes more sense than
reading mundane manuals :-)

Actually, would there be a big performance boost when using mode4?
Currently I am seeing 95% total throughput. Which isn't that bad. I am
peaking at 238MB/sec (each gig/e connections)

Also, mode0 does fault tolerance, meaning if a switch failure occurs
we should still be good, but how would the packets then be
transferred? I suppose rr would be disabled since it won't need to
alternate, correct?

On Sat, Sep 6, 2008 at 12:15 PM, Filipe Brandenburger
<filbranden at gmail.com> wrote:
> On Sat, Sep 6, 2008 at 11:57, Mag Gam <magawake at gmail.com> wrote:
>> Suppose data is being pushed out, it will go with 1st NIC and when it
>> gets overloaded it will use 2nd NIC.
> No. If you are using "balance-rr", one packet will go through the 1st
> NIC, and the next packet will go through the 2nd one. That's what "rr"
> (round-robin) means.
>> Similar to the push, the pull will be very similar. The data gets
>> pulled and the bonding driver will assemble the packets together? Does
>> this sound right?
> Actually this will not be determined by the bonding driver, it will be
> determined by the switch that is actually "pushing" the packets. The
> bonding driver will only make it look like the packets are coming from
> one (bonded) interface only.
> How the switch will behave depends on its configuration. It may be
> configured to send all the data through one of the interfaces only to
> balance through both of them using round-robin or something else.
> You should try to read this, it's very complete:
> /usr/share/doc/iputils-*/README.bonding
> Also, if your switch supports it, you should try to use the 802.3ad
> mode (mode=4) since that will probably give you the best results with
> bonding (in terms of load balancing and fault tolerance).
> HTH,
> Filipe
> _______________________________________________
> CentOS mailing list
> CentOS at centos.org
> http://lists.centos.org/mailman/listinfo/centos