网络课程笔记

00:00 Hello and welcome to this lecture.

00:01 So in the earlier lecture we saw enhanced networking and then how to enable enhanced networking for

00:06 your EC2 instances.

00:08 Now in this lecture we are going to see some additional network optimization techniques.

00:12 Now this hasn't been seen in the exam much, but this is something good to know because you may expect

00:18 some questions sometime on this.

00:20 So there are two additional techniques that we are going to see.

00:23 One is depdc and another is using the EFA.

00:27 Now let's see what that means.

00:29 So Depdc is a Intel data plane development kit.

00:32 And this is nothing but a set of libraries and drivers for fast packet processing.

00:37 Now what we are now talking about is not really about while the traffic flows from one system to another

00:43 system, um, via the hypervisor and then, uh, through the drivers or a Nic, what we are talking

00:50 about processing those packet packet fast before sending it through the network.

00:55 And that's what the Depdc does.

00:57 It basically, um, increases the performance of processing the packet.

01:02 Now while enhanced networking and sr-iov reduces the overhead of packet processing between instance

01:07 and the hypervisor, as we just understood in the last lecture, Depdc reduces overhead of packet processing

01:15 inside the operating system.

01:17 That means you can process much more packets than usually you can if you use Depdc libraries in your

01:24 operating system.

01:26 So Depdc ultimately provides lower latency due to kernel bypass.

01:30 And what does that mean we will see shortly.

01:33 It gives you more control of packet processing, and it ultimately lowers your CPU overhead for processing

01:39 the network packets.

01:41 That's what Depdc does.

01:42 So whenever the question is about enhancing the performance within the operating system, then it is

01:48 more about depdc.

01:50 Now this is a typical flow.

01:52 When the traffic flows from your one VM to another, there is an application which is sending a packet

01:56 to another machine and another application.

01:59 So packet obviously flows through your kernel and then it goes through virtual function driver either

02:05 through Ina or Intel virtual function.

02:07 And then the network will flow and the receiving end VM will receive that packet.

02:12 Now which means this packet is being also processed by your operating system.

02:16 Or you can call it a kernel.

02:17 Right.

02:18 Now what Depdc allows you to do is to bypass this kernel altogether.

02:23 If you use Depdc enabled libraries, packet can directly go from your application to directly your,

02:31 uh, driver which is attached to your EC2 instance.

02:34 And that's why you can process now much more packets than you would expect otherwise.

02:39 So that's in short, is depdc.

02:41 There is nothing much to discuss here, but just remember that it's an it's an operating system level

02:47 optimization for processing a maximum number of packets per second.

02:52 Um, there is one more concept that you need to know.

02:55 Um, other than Depdc.

02:57 As I said, there is something called EFA which is elastic fabric adapter.

03:01 Now we we talked about uh, Ina in the last lecture, which is, which is a network adapter.

03:07 And when you attach a network adapter to your EC2 instance, it can give you up to 100 GB s bandwidth.

03:14 Uh, but elastic fabric adapter EFA is a special type of Ina with some added capabilities basically.

03:22 And the limitation is it can be it can be only used with Linux machine.

03:26 That means if you attach EFA to windows machine, it would very well act as a Ina only.

03:33 So remember that now EFA further enhances your networking capabilities.

03:38 You need to then for that understand what is MPI which is message passing interface.

03:44 And what it does is basically it's a communication protocol for parallel programming.

03:49 So whenever you talk about HPC workloads, you know that a same compute workload can be run into multiple

03:55 machine in a distributed fashion.

03:57 And in that case, MPI will allow the parallel communication between all these nodes and ultimately

04:04 increasing your network throughput.

04:06 So there is nothing much to discuss about the EFA at the moment, but understand that EFA is just and

04:14 enhance or Ina with added capabilities, and it is applicable for Linux machine and not for the windows

04:21 machine.

04:22 So with that, I think this is a short um, session where we just talk about two things.

04:27 First is Depdc, uh, and second is EFA, which can further enhance your network throughput or bandwidth.

04:35 So thanks for watching.