Across the DC space in general most providers use NVO3 and vxlan source
port entropy L2/L3/L4 hash which provides per packet uniform 50/50 load
balancing at the L2 VNI overlay layer, which translates into underlay load
balancing of flows and thus no polarization.
Across the DC space speaking from
Hi Arie,
Draft draft-ietf-idr-link-bandwidth talks about advertising towards IBGP.
It does not talk about advertising over EBGP.
While I do support your use case I think it would be much cleaner to just
ask for new ext. community type.
Reason being that as you illustrate you may want to
Jeff,
Actually, the way this draft is written, and how the implementations I'm
aware of are implemented, this is not really a transitive community. It is
a new community that is being generated on the AS boundary.
The community value is not carried over, but is calculated based on an
cumulative
Hi,
I support adoption of the draft as Informational, please note, that request to
change transitivity characteristics of the community is requested in another
draft.
Gyan - please note, while pretty much every vendor has implemented the
community and relevant data-plane constructs, initial
Hi Robert
Very good point on millions of mice flows that can take advantage of
unequal cost lb, versus a small number of elephant flows.
Good point on further granularity with TE bandwidth management .
Gyan
On Tue, May 25, 2021 at 2:06 PM Robert Raszuk wrote:
> Gyan,
>
> It is always
Arie
Thanks for responding on the polarization question and I agree it can
enhance ECMP capability and maybe even counter or reduce effects of
polarization.
Thanks
Gyan
On Tue, May 25, 2021 at 1:26 PM Arie Vayner wrote:
> The flow polarization or elephant flow issues are well known industry
That’s great news that Cisco had implemented and customers have deployed
for years!
I see it’s supported in IOS and XR
https://www.cisco.com/c/en/us/td/docs/ios-xml/ios/iproute_bgp/configuration/xe-16/irg-xe-16-book/bgp-link-bandwidth.html
Gyan,
It is always helpful to an assessment into right scale.
Yes if you take few flows perhaps even a few big ones may suffer from
polarization. But the feature here is about hashing millions of micro
flows. With that in mind polarization effects are insignificant at
decent operational scale.
The link bandwidth community has been implemented by Cisco and deployed by
our customers for several years.
Polarization of flows in multipath is a well known problem, but it hasn't
deterred
people from using it.
Regards,
Jakob.
From: BESS On Behalf Of Gyan Mishra
Sent: Tuesday, May 25, 2021
Hi Satya
I read the draft and have a few questions.
IPv4 does not support per flow per packet load balancing as all packets
belonging to the same flow must hash to the same path to prevent out of
order packets and thus is subject to polarization of flows as high
bandwidth flows may hash to the
10 matches
Mail list logo