Building AI Data Centers with Apstra - Optimizing Fabrics for Best Performance with DLB (Demo)

0:00 For this demo.

0:01 Let's talk about ways that Apstra can help optimize

0:04 A.I. fabrics for the best possible performance.

0:07 We'll use our representative customer, A1 Arcade, as our example.

0:12 As you know, load balancing is an important way to optimize link capacity.

0:16 But Juniper's solution goes a step further using dynamic load balancing,

0:21 which takes into account link utilization when hashing flows.

0:24 For more intelligent traffic distribution. Configs like load balancing

0:28 would normally be configured upfront,

0:30 but I'm going to show you how easily Apstra can add configurations

0:34 to A1 Arcade's fabric in a live network scenario.

0:37 While we have our rail optimized fabric which can reduce spine traffic,

0:41 we still need to optimize for efficiency and performance.

0:44 This is where Apstra can really help. From the dashboard Apstra

0:49 analytics shows us that A1 Arcade's link utilization is unbalanced.

0:53 You can drill down to look at byte and packet stats,

0:56 but ultimately we're going to need to add DLB.

0:59 To do this, I can just go to our Apstra library and pull down a DLB configlet.

1:04 Once staged, we simply commit the changes to the fabric

1:08 and we've resolved the bottleneck.

1:10 You can look at historic performance over time to show before

1:14 and after profiles to measure the impact of our efforts.

1:20 Shifting to the front end.

1:22 A1 Arcade needs to use technologies like explicit congestion

1:25 notification or ECN and priority flow control or PFC to shut down

1:31 traffic congestion both end to end and within the fabric respectively.

1:35 During periods of high congestion.

1:37 ECN uses congestion notifications in the packets to reduce the transmission

1:42 rates of the servers transmitting traffic into the fabric.

1:45 This shuts down the congestion at the source to provide immediate relief.

1:50 PFC targets the switch fabric using congestion notifications.

1:54 This allows the switch that is sending traffic into a congested fabric

1:58 to throttle low priority traffic without impacting the throughput of high

2:02 priority traffic.

2:04 Back at the Apstra dashboard.

2:05 We are alerted to a spike in congestion at the spine.

2:09 Like our previous example with DLB, Apstra makes configuration changes

2:13 to a live network easy and instantaneous

2:16 to update the fabric and correct the congestion.

2:19 We simply pull down a cost configlet from the library and commit it to

2:23 the network.