Part 3 of 3 Video Series: IP Fabrics for Modern Data Centers 

00:00 [Music]

00:12 thanks for joining me for the last

00:13 session on the dc architectures

00:16 in the first session we discussed the

00:18 bgp unnumbered standard specified in rfc

00:22 5549

00:24 becoming increasingly popular in the

00:26 data center ip fabric ecosystem

00:29 how it compares to ebgp and how it is

00:32 easier to enable compared to the igp's

00:35 in the second session we discussed rocky

00:37 v2 for boosting data center efficiency

00:41 and data delivery performance

00:43 michael thanks for joining me for our

00:45 last session it in these has been a

00:47 pleasure to discuss foundational topics

00:50 and questions

00:52 as they are on top of the mind from many

00:54 data center technologies hey aaron

00:56 thanks for inviting me again it's a

00:58 pleasure as well on my end

01:00 thanks michael

01:01 so today michael i would like to double

01:03 click into another fundamental topic of

01:06 why ip fabrics using the the three stage

01:10 or the five stage data center

01:11 architectures are becoming the de facto

01:14 standard for modern data centers

01:17 so the question to you is uh michael a

01:20 traditional land data center used to be

01:22 deployed using access aggregation core

01:25 legacy architectures

01:27 and some level of ip awareness was also

01:29 there so why the ib fabric or ip class

01:33 are becoming so popular yeah good point

01:35 aaron so as a matter of fact so ip

01:37 fabrics are around for uh the last

01:39 couple of years actually highly deployed

01:42 by uh the main cloud providers we

01:45 observed that they got adoption also in

01:47 case of enterprise and telco clouds

01:49 right because of the robustness right so

01:52 we don't use any more of things like a

01:54 spanning tree with active standby

01:56 approach in case of ip fabrics we simply

01:59 use the active active

02:01 ipec and p forwarding from the the leaf

02:04 to spine and super spine or from t0 t1

02:08 and t2 a block of architecture right so

02:11 we are not relying on any proprietary

02:13 mechanisms anymore and it's quite often

02:15 based on ebgb simply uh at the leaf at

02:19 the spine and super spine and we get a

02:23 benefit of open standard architecture

02:25 which is

02:26 defined in some of the rfc documents for

02:28 example the one that is

02:30 7938

02:32 is is precisely describing what are the

02:35 recommendations to deploy that type of

02:37 ip fabrics highly scalable data center

02:40 bgpip fabrics right

02:42 excellent that makes a lot more sense so

02:46 what are the main physical topology

02:48 differences for ib fabric versus the

02:51 traditional layer 2 layer 3 ethernet

02:54 access core aggregation so in fact uh

02:57 aaron so when when you look at the three

02:59 stage five stage

03:00 type of architectures you don't have

03:02 usually the back-to-back links at the

03:05 given

03:06 stage of the architecture so for example

03:08 the tor or leaf device is usually not

03:11 connected to another tor device right

03:14 so in the legacy architecture sometimes

03:17 it was the case in case of ip fabrics we

03:20 don't connect at the given stage of the

03:22 architectural level of the architecture

03:25 by back-to-back connect

03:27 we don't use that anymore and when it

03:30 comes to the connectivities we don't

03:32 rely on any active standby model anymore

03:34 everything is uh just active active

03:36 based on the fact that we can use uh ipe

03:39 cnp on the chip itself right to reach

03:42 the destination ip we can go through

03:45 multiple links from the tour to the

03:47 studio to the to the spine right

03:50 and then uh obviously we can easily add

03:52 links into the architecture right so i

03:55 already mentioned that during our

03:57 storage video

03:58 and uh simply whenever the data is

04:00 growing when you're in in the data

04:02 center we can easily add these links to

04:05 the existing infrastructure without

04:08 changing the the design

04:10 dramatically right

04:12 and then the last thing is that the l2

04:14 domain

04:16 diameter is uh actually reduced or

04:19 contained in the top of the rack or

04:22 eventually in the pair of top of the

04:25 racks right so we don't extend these l2

04:27 domains usually in case of native ip

04:30 fabrics it's contained in the top of the

04:32 rock the rest is just ip routing right

04:34 using bgp or any of the igps we have on

04:37 the market right so these are the main

04:40 differences from the physical point of

04:42 view that you know we just follow the ip

04:45 clause type of architectures uh that is

04:48 predominantly used

04:50 in case of uh cloud provider networks

04:53 right so

04:54 we talked about the physical

04:55 characteristics of an ipv fabric what

04:58 about the logical part of the of the

05:01 design

05:03 hey so from logical point of view uh

05:06 traditionally we had in the access core

05:08 aggregation aggregation core

05:10 we had the uh first of ip gateways

05:12 somewhere in the aggregation or

05:14 sometimes even in the core usually in

05:16 the aggregation in case of ip fabrics

05:18 the first pipe gateways usually are

05:21 either the tor at the leaf for device

05:23 level right so that we can deploy on any

05:26 cast type of model and then uh all the

05:29 servers can uh can reach this personal

05:31 pipe gateway directly at the top of the

05:34 rack and so it means automatically that

05:38 that we reduce the blast radius in case

05:40 of any upgrades or failures of of the

05:43 top of the rocks

05:44 and then the other thing that is more

05:46 specific to pure ip fabrics is that

05:48 we have also a trend where actually the

05:51 server runs bgp

05:53 or any other form of routing protocol

05:56 and just connects through bgp to the top

05:58 of the rack

06:00 and then at the server level we just

06:02 through that bgp we advertise some uh

06:05 sort of of a veep

06:07 type of prefix

06:08 from all different servers so servers

06:10 are running bgp they connect through vgp

06:13 to the to the top of the rack that's the

06:15 new trend we also observe in the tier 2

06:18 cloud providers

06:21 and it's quite interesting because

06:23 thanks to that approach we are just

06:26 reducing the uh the utilization of the

06:28 lacp which is sometimes an easy thing to

06:32 put in place on the server but sometimes

06:34 it's not as easy as it uh as it as as it

06:37 is on the paper right and then the last

06:39 point is uh the load balancing

06:41 capabilities so whenever we start

06:43 investing in 100 gig 400 gig 800 gig we

06:47 want to make sure that the load

06:50 balancing capabilities of the top of the

06:52 rack are uh efficiently uh implemented

06:55 right so that this bandwidth is not

06:58 wasted right so that we use all of the

07:01 possible links in order to get the most

07:03 of the performances from the network

07:06 amazing absolutely incredible uh michael

07:09 so i appreciate the the discussion i

07:12 also appreciate

07:14 your time because these discussions have

07:16 been remarkable and very educative and

07:18 i'm sure i've learned quite a bit and

07:21 i'm sure our viewers too

07:23 uh so for all our viewers uh with this

07:25 we conclude our three-part video series

07:28 please feel free to reach out to juniper

07:30 networks for additional information or

07:32 questions

07:33 thank you

07:35 [Music]