网络课程笔记

00:00 Hello, and welcome to this lecture.

00:01 Here, we are going to summarize what we learned

00:04 about optimizing the network performance

00:07 while operating in AWS,

00:09 specifically with respect to the EC2 instances and VPC.

00:13 So let's understand this and revise this concept.

00:17 So for high network bandwidth and throughput,

00:20 consider using Jumbo Frames,

00:21 that is the first thing we saw.

00:22 With Jumbo Frames, you can send 9001 byte

00:27 in one single packet.

00:29 That means you can send more data

00:30 by sending the less number of packets.

00:34 That's the first thing.

00:35 Then you can enable the enhanced networking

00:37 for your EC2 instances by various mechanism that we saw,

00:42 you can use cluster placement groups,

00:44 so that EC2 instances are in close proximity,

00:47 and then they can get benefited of that

00:50 and send more packets and with lower latency,

00:53 then you can also use EBS optimized instances,

00:56 so that you get a dedicated bandwidth

00:58 for a communication between EC2 an EBS volume,

01:01 and then a separate bandwidth

01:03 to get for other network activities.

01:06 And for OS-level optimization,

01:08 you can use DPDK libraries,

01:10 that's Intel Data Plane Development Kit,

01:14 and with that, there is a direct communication

01:16 between your application and then it can directly send

01:21 the data to underlying network cards,

01:24 via those drivers, whether Intel Virtual Functions or ENA.

01:28 So it runs in a user space

01:31 and it bypasses the kernel when you use DPDK.

01:35 So just going over it again.

01:37 So instance level network optimization

01:39 can be done with enhanced networking

01:40 and for enhanced networking,

01:41 you have to enable the SR-IOV,

01:43 that is single route input-output virtualization.

01:46 With PCI pass through,

01:48 what it does is, ultimately it presents

01:50 your virtual physical network card

01:53 as a virtual network card to your instances,

01:56 and same physical network card can be shown as,

01:59 or can be represented as multiple virtual instances,

02:02 and then it bypasses the hypervisor

02:05 for the network communication.

02:07 Then you can always use ENA or Intel Virtual Function

02:10 in order to enable the enhanced networking.

02:13 You can also use placement groups as you know,

02:16 specifically cluster placement group,

02:18 where those EC2 instances will be typically launched

02:20 in a same availability zone in the same rack,

02:23 so that you get the lowest latency possible

02:26 between these two EC2 instances or multiple EC2 instances.

02:29 And finally, you can use EBS-optimized instances as well.

02:33 For operating system level network optimization, use DPDK,

02:36 we just talked about that.

02:38 And you can additionally use EFA.

02:41 EFA is nothing but ENA with additional capability

02:46 where you can bypass the OS

02:48 and that improves your network performance

02:51 typically would be used for HPC workloads,

02:54 HPC is your high performance computing workloads

02:58 like a media processing or telecom providers

03:00 typically would require HPC kind of workloads.

03:03 So that's in short summary of what we saw in this lecture.

03:07 If something that is not clear,

03:09 go back and you can watch the video again.

03:11 So thanks for watching.