Don’t Let Your AI Get Caught in Traffic

0:00 don't get caught in traffic don't play

0:02 in traffic and stay in your lane today

0:05 we're going to be paying attention to

0:07 exactly none of that and doing the exact

0:10 opposite I'm James Kelly from Juniper

0:12 Networks we're going to be talking about

0:14 playing in traffic at high speeds in AI

0:17 clusters and exactly how you manage that

0:19 let's get into it all right let's take a

0:21 look at our handy terraform abstra

0:24 examples repository if you were with me

0:27 in the last demo about AI topology

0:31 design you'll know that this demo is

0:33 going to show you how to automate abstra

0:36 with our abstra terraform provider for

0:39 AI clusters last time we were talking

0:41 about rail optimized design an Nvidia

0:43 prescribed feature and how we make that

0:45 easy with juniper abstruct in our intent

0:48 based multivendor fabric manager at

0:50 Juniper Networks today we're going to be

0:52 talking about a different part of our

0:56 best practice recommended design and

0:58 last time we were looking at this

0:59 subfolder of the examples I'm going to

1:01 go into this example lab of a real AI

1:06 cluster that we've set up with three

1:08 different Fabrics or three different

1:10 so-called blueprints in abro and I've

1:15 actually got an instance of apure Spun

1:17 up in cloud Labs you'll know that I

1:19 mentioned last time too that uh you can

1:21 delete and create your own topologies

1:23 here and uh just at a few minutes after

1:28 a click of a button you'll have access

1:29 access to abstra and be able to log in

1:32 so when you open it up in a new tab you

1:34 basically have this and it asks you to

1:36 log in and it gives you the credentials

1:37 right

1:38 there also like last time I'm going to

1:42 start off by uh showing you my GitHub

1:45 desktop I've just pulled down the

1:48 repository that I was showing you in my

1:49 browser terraform apture examples that

1:52 has all of the Juniper examples in there

1:56 and we're going to be using that one of

1:57 them of course so I'm going to open this

1:59 up up in Visual Studio

2:02 code like last time the first thing that

2:04 I have to do is actually point my

2:07 provider at my instance of

2:12 abstra and I'm going to just copy and

2:15 paste this I had a quick note outside of

2:18 the screen recording that's how I did

2:20 that so easily but admin is the username

2:23 amazing catv your dollar sign is the

2:25 password and this is the IP address that

2:27 you saw in the last

2:30 browser tab let me just flash back there

2:33 so

2:34 that you know you can see exactly what

2:37 I'm talking about here's the IP address

2:40 here's the credentials all right

2:44 so back to this I'm just going to now uh

2:49 save this

2:50 file and take a look at uh what we're

2:54 talking about in our blog was Dynamic

2:57 load balancing and how it helps

2:59 alleviate congestion over regular ecmp

3:03 load balancing and the possibilities of

3:06 using some packet spraying in the future

3:08 and other forms of load balancing um as

3:11 part of abstra reference designs those

3:14 intent-based very nicely validated

3:17 designs that are laid down as a data

3:20 center Network fabric um we're going to

3:23 be using the layer 3 routed only designs

3:27 and as part of that it doesn't include

3:30 the configuration in junos for dynamic

3:33 load balancing if you want to learn more

3:35 about Dynamic load balancing I'm going

3:37 to keep the video short and ask you to

3:39 go and read the details in the blog it's

3:41 also a little bit of self promotion

3:42 since I wrote the blog I would encourage

3:45 you to read it I would hope um this is

3:47 the actual very simple configuration in

3:49 junos and if you're familiar with junos

3:51 this should the stanza sort of syntax

3:54 look very familiar to you you can change

3:57 the inactivity timer in microsc or you

4:00 can just leave it out completely and

4:01 it'll default to

4:03 256 I wanted to show that to you um also

4:07 unlike last time where we only looked at

4:09 designs I've got a blueprints file here

4:12 and this really creates the blueprints

4:15 creates some of the resources that are

4:16 necessary for the blueprints and I'm

4:18 going to be applying as I described in

4:21 the blog the configlet that was created

4:25 in that other file to two different

4:27 Fabrics one for my back end GPU to GPU

4:31 Fabric and one for my storage fabric

4:34 both of those are the Rocky very high

4:36 bandwidth fairly low flow count Fabrics

4:40 that'll highly benefit from Dynamic load

4:42 balancing as compared to you know other

4:44 kinds of load balancing and as an

4:47 example you could apply this to all

4:48 devices we've just you know given this

4:50 little condition in here if you only

4:52 wanted to apply it to the leaves for

4:54 example if you had let's say single

4:56 links between your leaves and spines

4:59 dynamic balancing really wouldn't help

5:01 you at all on the way back from The

5:03 Spine down to the leaf that would be an

5:05 example of how you would do that all

5:07 right so I'm not going to go into any

5:09 more about the configlet I'll let you

5:11 check that out at your own Leisure what

5:13 I will show you though is that coming in

5:17 here like last time we can before I do

5:20 this I need to change into the right

5:25 folder and I'll do a terraform in it we

5:29 just make sure that I have the right

5:31 version installed the most recent

5:32 version of the abstra terraform

5:36 provider and if I do a terraform plan

5:39 this is a know blank slate in terms of

5:42 my instance of Abra that I spun up it'll

5:44 say that there's 58 different resources

5:47 to

5:48 add and I will do a terraform

5:52 apply and before I type yes what I can

5:56 do I suppose is perhaps this just kind

5:58 of split the the screen here and let's

6:01 look at

6:03 abstra and you'll remember how all of

6:05 the designs showed up for example under

6:08 the racks and the templates in the last

6:10 demo that I did this time if we go into

6:13 blueprints you see that there's

6:14 absolutely no blueprints here right

6:18 now as soon as I say yes here it's going

6:21 to start firing away at the abstra AP

6:25 the teror form provider for Abra is

6:27 implemented in go and that uses a go SDK

6:30 for abstra was also um open sourced and

6:34 all of this stuff is going just in a

6:36 matter of seconds creating three

6:39 different fabric configurations inside

6:42 of abstr now what this terraform

6:45 actually doesn't do that will be adding

6:47 in the future as part of my next demo as

6:49 I'm going step by step through this is

6:52 additional configlets I talked in the

6:54 blog about you know hinting at bcq CN in

6:56 the next and then looking at some

6:58 Analytics um so this isn't actually

7:01 going to commit anything to the Juniper

7:03 Juno devices we'll save that for a later

7:05 demo but you can see basically now that

7:07 the terraform stuff is all done it's

7:10 gone and created the backend GPU fabric

7:13 backend storage Fabric and let's now

7:18 make this window a bit

7:19 bigger the front end management Fabric

7:22 and if you look inside of any one of

7:24 these you'll have to come into the stage

7:27 Tab and you can see that there's certain

7:29 types of resources that were allocated

7:31 this was done in that blueprints uh. TF

7:34 file and there's certain things for

7:36 example like the spine and the leaf that

7:39 aren't yet actually allocated right so

7:43 if you wanted to for example go and you

7:46 know allocate those here you could do

7:48 that manually as I said we're going to

7:49 save all of that automation for a later

7:52 demo and show that to you as well and of

7:55 course none of this is committed because

7:57 for it to be committed we have to be

7:58 using real devices as I'm putting this

8:00 together quickly uh like I said one step

8:02 at a time so we will get there in later

8:04 demos stay tuned for how this actually

8:07 shows up at the device level and then

8:10 you can kind of see for example when you

8:12 see one of these Juniper devices all of

8:14 the configuration for it and we'll take

8:17 a look at how that configlet gets

8:18 applied here in the next demo in the

8:21 meantime you can see that uh this is the

8:23 front-end fabric where we actually

8:24 didn't apply the config so I should look

8:26 at let's say the storage or the other

8:28 one and when I go in there if I go into

8:32 the config lates here excuse me you'll

8:34 see this config here DLB for AI leaves

8:39 that was the config that we provide uh

8:42 provisioned excuse me using

8:46 terraform all right that concludes the

8:48 demo I hope you'll go and check out this

8:50 repository you can easily pull it down

8:52 and walk through these things using

8:54 Astra Cloud labs and you can see how

8:56 easy it is to create and automate stuff

8:58 in asra

9:00 and stay tuned for the next step on the

9:03 journey to creating and automating AI

9:06 training clusters with Juniper Networks

9:09 I'm James Kelly thank you