0:00

The. Two

0:02

and Half Edmunds Episode One, Eight, Five, I'm.

0:04

Joe. I'm. Jim and I'm Helen.

0:06

A hero again. Last. Time

0:08

you talked about privacy. This. Time Let's

0:11

talk about ai. Jen. Hsun Huang

0:13

says kids shouldn't learn to code. They

0:15

should leave it up to a I

0:17

G. What? A strange thing for

0:19

a guy running in video to say

0:21

moon yes Not like they have a

0:23

vested interest in selling hardware to people

0:26

who do a i bullshit or anything

0:28

gym at the same time. Doesn't.

0:30

And video need to be able to

0:32

hire people to write the Ai stuff.

0:35

Know they I'll do the math. presumably.

0:38

Wang is trying to talk to the kids who

0:40

aren't going to grow up to be the kind

0:42

of engineers that in video needs or wants because

0:45

a I certainly isn't doing that kind of job

0:47

any time soon. presumably. He

0:49

just figure sees can I'm. Screwing.

0:52

Over the next generation from I guess

0:54

lower and coders. And I just

0:56

just turning over low to mid range coating

0:58

the ai. But. You

1:00

can't rely on a I for that either. Yeah,

1:02

you can rely on it to help you out

1:05

save you some time, but you still need the

1:07

skills to understand what spitting out a. You.

1:09

Can rely on it to give you ideas

1:11

is what you can rely on Iran, but

1:14

whatever ideas give you, they need to be

1:16

thought of that way. Ideas untested, unproven that

1:18

you need to check, and in many cases

1:20

that can still be. You know, a huge

1:22

efficiency when productivity when an Ai tool can

1:25

break you out of just like a brain

1:27

lock where you can't think of the thing

1:29

you need to think of and you need

1:31

some new data. Just kind of. Job.

1:34

Degree Meet. Sometimes A I can help

1:36

you by producing a whole bunch of

1:38

bull or break crap that doesn't change much

1:40

from one thing to the next and

1:42

it's like a huge mental or emotional

1:44

or energy drain for you to sit

1:46

there and type it all out. But if

1:49

it just gets morphed out in front

1:51

of you, your competent to scan it

1:53

in see this is correct or not

1:55

correct. What you really can't do with a

1:57

I and Co to say hey I.

2:00

write this code for me and I'm just

2:02

going to hit save and run it. Yeah,

2:04

I'm specifically talking about boilerplate stuff that like

2:06

you say, it's just typing out loads of

2:08

stuff and it just goes blah instantly saves

2:11

you 30 seconds. But

2:13

again, you really can't use it for that

2:16

unless you understand the boilerplate stuff because it

2:18

won't always get that right either. Yeah. So

2:21

if you can't catch and correct an error

2:23

in the boilerplate stuff faster than you can

2:25

type the boilerplate stuff out, it's still kind

2:27

of a lose. Nobody

2:30

in the world isn't going to be a

2:32

coder and every single kid isn't going to

2:34

either. But

2:36

I still think there's value in them understanding some of the

2:38

concepts of it. And I wish

2:40

I had got introduced to the concepts in

2:42

a more formal way earlier in my career

2:45

rather than I learned to program

2:47

by downloading QBasic code from the

2:49

local BBS and reading it and

2:52

muddling with it and learning the hard

2:54

way. And then by the

2:56

time they started trying to teach the basics of

2:58

programming in high school, it's like I was running

3:00

circles around that. Yeah, and you accidentally

3:02

became a developer because you were a sysadmin

3:05

originally and then you sort of just accidentally

3:07

became a developer. Whereas I thought I was

3:09

going to be a software developer my whole

3:11

life and then discovered no, actually turns out

3:13

I'm a sysadmin. Yeah, for me it

3:15

was a little bit like I thought I was going to be

3:17

a software developer at first as well. But

3:19

then in high school, I did a job placement at

3:22

the power plant as a sysadmin and I really enjoyed that. And

3:25

then when it was time to decide what to

3:27

go to school for, it's like, well, all the

3:29

programming I know I taught myself or I can

3:31

read from a book and I can do with

3:33

my one computer. But to be a sysadmin, I

3:35

need a lot of computers. And

3:38

guess what the college has? A whole lot of computers

3:40

I can go in touch with. So

3:42

that's what I'm going to school for because that's

3:44

where I'm going to actually get more value from

3:46

the practical side of it. Alan, you

3:48

are pretty pragmatic. Have you tried

3:50

GitHub CodePod for example? Not

3:53

the CodePod one. I've used chat GPT to

3:55

help come up with ideas for articles to

3:57

write or how to structure a presentation or

4:00

how to just say something or questions

4:02

to ask during job interviews and lots of

4:05

ideas, like Jim was saying, asking it, here's

4:07

some things I've thought of, think

4:09

of 10 more for me to help me fill out

4:12

this list. It's reasonably good at

4:14

that. You need to also

4:16

apply a critical eye and edit it in a bit

4:18

and so on, but it can really be a time-saver.

4:21

But for source code, it's like with

4:23

what I'm doing in the source code, honestly,

4:27

probably not going to be as helpful or my

4:29

biggest concern still is, what is the license

4:32

of this code going to be? Because

4:34

if I'm selling this code to people, I

4:36

have to be able to rightfully

4:38

claim that I wrote this and

4:41

that I have the authority to sell

4:43

it to you and that it isn't

4:45

some other code. As far as

4:48

Wang's idea of saying that kids should

4:50

learn AI instead of coding or whatever,

4:52

I think the big thing here that keeps

4:55

getting me about that is the issues of

4:58

who wrote it, who owns it, is

5:00

it clean in terms of the

5:02

bill of materials, whatever for

5:04

AI written code. Even when you put all that

5:06

stuff off to the side, just the idea that

5:09

do it with AI instead of doing it yourself,

5:12

I have the same advice there that I give people

5:14

all the time about scripting. I see larval

5:17

folks all the time writing these really

5:19

long scripts to chain together 20 or 30 simple tasks

5:22

that they don't really understand the individual tasks very

5:24

well yet. They may very

5:26

well be able to put together a script that ties

5:28

all those 20 or 30 things into one script that

5:30

they can run to do that thing over and over

5:33

again, but you

5:35

haven't learned how to actually do

5:38

any of it. If you lose that script,

5:40

you're screwed. There's a lot of things you

5:42

don't understand. My advice is do

5:45

those 20 or 30 things manually, frequently enough,

5:47

often enough that you've got no problem just

5:49

doing it all off the top of your

5:51

head, then write the script because you

5:53

don't need to do those things anymore because you know them,

5:55

you understand them, you know all the steps, you can do

5:58

them without the script. look

6:00

at, okay let me take the script and every

6:19

time almost certainly won't

6:21

do the same

6:24

thing every time. It is partially random

6:27

in the same way that you know a dog

6:29

is partially random. That's the whole

6:31

point. You didn't have to program it. You

6:33

could just give it a loose idea of

6:35

what you wanted it to do in

6:37

a sloppy, imprecise human

6:40

language and it would mostly do what

6:42

you mostly wanted. So it's like

6:44

anything else. When that's the degree of accuracy

6:46

you have from the person on the top

6:48

to the person next down the chain stretching

6:51

our terms a little bit and granting

6:54

you know some large language model the

6:56

title person, still the point is it's

6:58

imprecise so you have to supervise in

7:01

a way that you don't as much with code

7:03

because you can find out whether the

7:05

code is buggy or not and you know

7:07

it'll operate the same way and if you're doing the same

7:09

thing it'll do the same thing and again with a model

7:12

not so much. But it may be

7:14

taken from a different lens if what

7:16

Jensen was trying to say or maybe

7:18

just a thought I'm having is rather

7:20

than every grade schooler learning

7:22

Python and only some of them

7:25

going on to actually do development if we

7:27

taught them how to use AI properly and

7:29

what its limitations are that might be much

7:31

more useful to a lot of them that

7:33

aren't going to become developers. The

7:36

skills of how to use AI to

7:38

get prompts for creative writing or art

7:40

or all the other things that AI

7:42

can be used for it might actually

7:44

be a more valuable skill than force

7:46

feeding Python to every grade schooler. See

7:48

what I hear you saying Alan is

7:50

kids should also learn to use AI

7:52

which I would agree with. What

7:55

Wang said is kids shouldn't learn

7:57

to code. Well yeah he says

7:59

it It is our job to create

8:01

computing technology such that nobody has to program,

8:04

and that the programming language is human. Everybody

8:07

in the world is now a programmer. This

8:09

is the miracle of artificial intelligence. And

8:11

it will be every bit as reliable as just

8:13

off-handedly telling your kids to go do something, and

8:15

then walking off and coming back four hours later

8:18

to see if it's done. Well, you

8:20

know what this reminds me of? It reminds me

8:22

of stuff like Dreamweaver back in the day, where

8:24

you could make a website, but anyone

8:26

who knew about making a website would look at

8:28

the code that Dreamweaver spat out and

8:30

go, ugh, this is horrible. I

8:32

mean, yeah, it kind of works,

8:35

but it's horrible code. Same thing

8:37

with Microsoft's front page. I

8:39

remember that and, like, you know, you bold something and then

8:41

change it and later it would be like an open

8:44

bold, closed bold tag around nothing and just all kinds

8:46

of detritus in the code. It was very useful as

8:48

a learning tool in that you could use the WYSIWY

8:50

to get something and then look at the source code

8:52

side and figure out how it did it. But

8:55

as soon as you understood anybody, you're like, why is

8:57

it doing all this other stupid stuff? I

9:00

used to use front page as a WYSIWYG editor

9:02

and I would actually write the code in HTML in

9:04

the source window, but just the fact that it

9:06

would render in real time as I was doing it

9:08

was invaluable because you didn't have a dev console

9:10

in the browser to do that with back then. Yeah,

9:13

I was like, I don't know how we ever did,

9:15

like, CSS without the developer tools in

9:17

the browser. With great difficulty and much

9:20

angst. Well, especially back then when it's

9:22

like, oh, you have to, like, upload

9:24

it somewhere before you can view it,

9:27

have it all work correctly. So you, like, edit

9:29

the file, save the file, copy and paste

9:31

over FTP to somewhere, and then go to

9:33

the browser and hit refresh for every single edit.

9:35

And the round trip time was terrible.

9:38

Especially when it was CGI bin stuff and, like,

9:40

you literally couldn't do anything with it on your

9:43

own computer. So you're editing, you know, these thousand

9:45

line code files on your own machine and FTPing

9:48

them up over dial up to serve from the other

9:50

end of the country and then hitting a web page

9:52

that would hopefully run your code and maybe put out

9:54

the thing that you wanted. Oh,

9:56

that sucked. God, that sucked a lot. Well,

10:00

what about Google cuts a deal with

10:02

Reddit for AI training data and also

10:05

automatic, the company that owns Tumblr and

10:07

WordPress is going to sell users data

10:09

to Mid-Journey and OpenAI. It's

10:11

disappointing, but again, it's not surprising, is

10:14

it? Not really, and I think we're going

10:16

to see more and more of this. The

10:18

kind of interesting thing is a lot of these

10:20

places probably didn't have a good cutout for

10:23

this in their terms of service. They

10:25

usually are generic as possible, so they can get

10:27

away with anything, but I think you

10:29

might start to see this specifically in terms of service

10:32

going forward saying, yeah, we could use

10:34

this data to train our own AI or sell it to

10:36

somebody else who will. But

10:38

I see a lot of places deciding

10:40

that, hey, this stuff we have, maybe

10:42

somebody will pay us for it and

10:45

going and doing that. And I think it's terrible,

10:48

but I also think unless we

10:50

make a really big stink really quickly, it's going

10:52

to happen a lot. I mean,

10:54

Jim, you wrote hundreds of thousands of words on

10:56

Reddit helping people out with a very set of

10:59

S problems and whatnot. You must

11:01

feel pretty bummed out that that's just

11:03

being sold to some AI company. This

11:06

may surprise you, but actually

11:08

that part doesn't bother me so much.

11:10

Well, I would say specifically you wrote

11:12

those things in public to be public

11:14

for the purpose of it being publicly

11:16

consumed, right? Exactly. I wrote those

11:18

things on a very public forum with the

11:20

idea that anyone who wanted to could learn

11:22

from it, and that includes the AI model.

11:25

Again, I will point out

11:27

that AI models learn. It's

11:30

not really that different from

11:32

the way non-artificial intelligence learns.

11:34

They train on data. I'm

11:37

not mad about somebody learning from

11:39

the things that I said online because that's my

11:41

whole goal is I want people to learn from

11:43

it. What I was mad about was the heavy-handed

11:46

actions of Reddit itself,

11:49

just screwing people over and

11:52

yanking the API away from a

11:54

third-party ecosystem that literally had to

11:57

grow up because Reddit itself wasn't

11:59

producing. using usable apps and

12:01

just the high handedness. The biggest

12:03

thing that got me as mad at Reddit as

12:05

I was was not even

12:07

yanking the API. It was when I

12:09

saw what Spiz had to say about

12:11

it. When I saw the leaked internal

12:14

memo that talked about, you know, oh,

12:16

well, you know, the user base is

12:18

noisy, but just rioted out. I was

12:20

like, Oh, oh, that's how it is,

12:22

huh? We're just noisy and inconvenient resources

12:24

that won't just sit there to be

12:26

quietly exploited however you want. We actually

12:28

insist on having some kind

12:30

of a voice. Hmm. Okay.

12:32

Yeah. I think it's also, it'll be

12:35

interesting to see the details if they ever

12:37

do come out about is

12:39

Reddit just providing a version of that

12:41

API where Google can in

12:43

a way that's less costly to Reddit scrape

12:45

all the text of people's posts, or

12:48

is this also including more metadata

12:50

like when they posted and how they post

12:52

and how people interacted with each post. If

12:55

it's just give us all the posts and

12:57

people made these posts in public Reddit, and

13:00

maybe it's not that big of a thing and Reddit does

13:03

have the right to that content. But

13:05

you know what it says, selling users

13:07

is data to train AI. That smacks

13:09

me slightly differently than just selling bulk

13:11

access to the content that Google could go

13:13

and scrape from Reddit. Normally, it would just

13:15

be much more efficient for Google and less

13:17

costly for Reddit if they use the API

13:20

to get it. I think ultimately, you

13:22

know, if this is something that we care about as

13:24

a society, you know, we have to actually

13:27

do something about it, whether the

13:29

we as consumers or the we as the

13:31

government that you know, the consumers elect, whether

13:34

the pushback is, you know, laws

13:36

that say you aren't allowed to do this, and

13:39

we will throw you in jail or make you

13:41

unable to do business in our country, or whether

13:43

the pushback is consumers refusing to buy or pay

13:45

for things, there has to

13:47

be actual pushback that matters or

13:49

companies aren't going to stop. The

13:52

phrase is don't leave money on the table.

13:55

And it's around longer than any of the three of

13:57

us have been alive or than any of our countries

13:59

have existed. That is human nature. You get

14:01

people in the business of making money, and they

14:03

want to make more of it, and they find

14:06

a way to do it. And unless

14:08

something slaps their peepee in

14:10

the door and makes them think that's a bad idea,

14:13

they will, because otherwise they're – and I'm doing the

14:15

scare quotes here – leaving money on

14:17

the table. Okay,

14:19

this episode is sponsored by TALIDE.

14:22

What do you call an endpoint security product

14:24

that works perfectly, what makes users miserable? A

14:26

failure. The old

14:28

approach to endpoint security is to lock

14:30

down employee devices and roll out changes

14:33

through forced restarts, but it just doesn't

14:35

work. IT is miserable because

14:37

they've got a mountain of support tickets, employees

14:40

start using personal devices to get their work done,

14:42

and executives opt out the first time it makes

14:44

them late for a meeting. You

14:46

can't have a successful security implementation unless

14:48

you work with end users. That's

14:51

where TALIDE comes in. Their

14:53

user-first device-trust solution notifies users as soon

14:55

as it detects an issue on their

14:57

device, and teaches them how to solve

14:59

it without needing help from IT. That

15:02

way, untrusted devices are blocked from

15:04

authenticating, but users don't stay blocked.

15:07

TALIDE is designed for companies with Okta,

15:09

and it works on macOS, Windows, Linux,

15:11

and mobile devices. So

15:14

if you have Okta, and you're looking for a device-trust

15:16

solution that respects your team, support the

15:18

show and go to collide.com/25a to

15:20

watch a demo and see how

15:22

it works. That's kolide.com/25a.

15:30

Let's do some free consulting then. For first, just a

15:32

quick thank you to everyone who supports us with PayPal

15:34

and Patreon. We really do appreciate that. If

15:37

you want to join those people, you can go

15:39

to 2.5admins.com/support, and remember that for various amounts

15:41

on Patreon you can get an advert-free RSS feed

15:43

of either just this show or all the

15:45

shows in the Late Night Linux family. And

15:48

if you want to send in your questions for Jibberlallen

15:50

or your feedback, you can email show at 2.5admins.com. David

15:54

writes, I'm considering an upgrade path for

15:56

my current router setup. At

15:58

the moment, I'm using a linksys.com. WRT1900ACS

16:01

router, which has a few 1Gbit

16:03

ports for WAN and LAN. I

16:06

don't have a 1Gbit internet connection, so on

16:08

that side, the speed of the 1Gbit router

16:10

is plenty enough. In a medium-term future,

16:12

I will be adding a NAS to my network, and

16:14

I'm wondering how I could upgrade the link speed to

16:17

2.5Gbits or more. If

16:19

I had a 2.5Gbit switch, and I connect most

16:22

of the wired devices directly to the switch, would

16:25

they be able to communicate at 2.5Gbits

16:27

between themselves, or will the router limit

16:29

the link speed to 1Gbit? The short answer

16:31

is yes, that's perfectly fine. As a matter of

16:33

fact, you're already doing that, whether you realize that

16:35

or not. Every gigabit switch out

16:37

there will also handle fast Ethernet at 100Mbit, or

16:39

the old 10Mbit Ethernet

16:42

even, all just fine. All the

16:45

devices on that switch will speak to each

16:47

other at the highest speed that their individual

16:49

link connection allows. The only

16:51

thing that I would warn you of here is,

16:53

and this is probably obvious, you've probably already thought

16:55

of this, it's not a real issue, but just

16:57

in case somebody hasn't figured this part out, if

17:00

you're thinking that you've got super fast

17:02

Wi-Fi clients, and they'll get a benefit

17:05

of that at 2.5Gbit, well, in

17:07

that case, then yes, your router or

17:09

access points had better also have 2.5Gbit

17:12

wired ports, or else you're not

17:14

going to see any of that theoretical speed up.

17:16

Not that they probably ever were going to go

17:19

anywhere near that fast anyway, because that's home marketing

17:21

hype. Yeah, you've got to just consider what is

17:23

connected to what, and if it's a 2.5 connected

17:26

directly to another 2.5, whether that's a

17:28

NIC in a desktop and a switch,

17:31

then that's going to work. But if

17:33

you have, say, a 2.5 port connected

17:36

to a 1Gbit switch that

17:38

is then connected to another machine that is

17:41

2.5, you're only going to get one. Yeah,

17:43

it's slightly more complicated. In a switch,

17:45

it's kind of virtually creating circuits between

17:48

every port to every port. And

17:50

so, yeah, all the ones that are linked at 2.5Gbits

17:52

will be able to talk at 2.5Gbits, although

17:55

in general, other traffic, like there's

17:57

a broadcast from one of the 1Gbit ports.

18:00

that's going to possibly live at the speed,

18:02

but only for microseconds and

18:04

so on. In general, it shouldn't

18:06

be a problem. Higher end switches

18:08

can do some amount of buffering to try to smooth

18:11

it out, but that's usually a, a

18:13

much more expensive switch than you're going to have, and b,

18:15

a complete waste of time at your house. My

18:18

house, I have a switch that has a

18:20

mix of 10 gigabit and 1 gigabit machines,

18:23

and my desktop and my NAS can

18:25

talk at 10 gigabit just fine, and

18:28

the other 1 gigabit clients can also talk

18:30

to the NAS just

18:33

fine. And you know, two

18:35

different 1 gigabit clients can talk to my

18:37

NAS at 1 gigabit each because

18:39

the NAS can deliver to the switch at

18:41

10 gigabits. With that said, you probably should

18:44

not expect to be able to feed 2 gigabit

18:46

clients at full wire rate from one single

18:49

2.5 gigabit connected

18:51

system. Alan might disagree

18:53

with me here, but I've only had limited

18:55

experience at 2.5 gigabit. It

18:57

has not impressed me. I think my

18:59

desktop has a 2.5 gigabit port, but it's

19:01

not hooked up to anything because I installed

19:03

a 10 gig NIC and ran it to

19:05

the other thing, so I don't have any

19:07

real experience with 2.5 gigabit either. Honestly,

19:10

most motherboards now sometimes

19:12

even have a 2.5 gigabit PHY, but

19:15

it's actually exposed out as just two

19:17

1 gigabit ports, which

19:19

is an even weirder configuration. So

19:21

yeah, I wouldn't expect to be able to do

19:23

that too much simultaneously, and it will really depend

19:25

on your switch how much you can actually

19:28

do the kind of multiplexing, but

19:31

you'll be totally fine to only upgrade some of your

19:33

devices in this case. Like you said, your internet connection

19:35

is only 1 gigabit or less than 1 gigabit, so

19:38

you don't have to replace your whole router

19:40

just because you want your desktop and your

19:42

NAS to be able to talk at more

19:44

than 1 gigabit. Absolutely. My recommendation here would

19:46

be absolutely play with this. Understand

19:48

that all you have to do is buy the switch. You don't have

19:51

to spend a ton of money and just... I

19:53

would advise you don't expect too much out of it,

19:55

which if you're not spending a ton of money is

19:57

not really a problem, right? Like you put the stuff

19:59

together, you see what it looks like. will do and

20:01

now you know and now you're a better admin or

20:03

hobbyist or whatever you like to call yourself because you

20:05

know more stuff. You've seen it happen. You know how

20:07

it works. And that's totally cool. What

20:09

I just don't want to see anybody doing

20:11

is being like, well, I think I can

20:13

scrape together the $400 or $500 for this

20:16

project that I expect to get these big

20:18

benefits out of. Like if that's the thought

20:20

process, I'm going to say maybe

20:22

don't do that with 2.5 gigabit because, again, I

20:25

don't have a ton of experience with it. But

20:27

I've seen it in action a couple of times

20:29

and not seen a whole

20:32

lot more than 1 gigabit of real

20:34

world throughput through it. And

20:36

the fact that you never ever see

20:39

2.5 gigabit stuff on anything

20:42

targeted for business, it's always like

20:44

home and gamer oriented. And in

20:46

business, you go straight from 1

20:48

gigabit to SFP plus and 10. There's no

20:50

in between there. Which I honestly found a

20:52

little bit weird because the whole point of the 2.5 gigabit

20:55

standard was take advantage of the 1

20:57

gigabit copper that's already in the walls

20:59

of your office and not have to

21:01

restructure your office. But I think it

21:03

– like I'm saying, maybe it's

21:06

just never delivered because I've not seen it

21:08

have any adoption on the business side. Well,

21:11

it's one of those things also. Like every time we

21:13

have a big bump up in network speeds, you never

21:15

get them in the first few years. Like you get

21:17

better than the last one and that's really what you

21:20

get. When 10 gigabit first dropped on

21:22

the scene, like nobody was getting 10 gigabit, they were getting

21:24

2 or 3 and they were happy about it. When

21:26

gigabit first dropped on the scene, nobody was getting

21:28

gigabit. They were getting like 500 or 600 megabit

21:31

and happy about it because again, that was 5

21:33

or 6 times the 100 megabit. As

21:35

the technology matures, you get closer and closer to

21:37

the theoretical wire rate in terms of the actual

21:40

wire rate throughput of data that you get. So

21:43

that now these days with gigabit, even

21:45

on just any random put together like

21:47

a couple of devices and a cheap

21:49

gigabit switch, you'll probably see 850 to 950 megabits

21:52

on an iPerF3 connection between the two. But

21:56

again, that's now. That's not what

21:59

it looked like. for the first five years or

22:01

so of Gigabit, you know, being out there in the

22:03

real world. I think that

22:05

2.5 Gigabit has kind of a similar problem

22:07

in that it's not really achieving, you know,

22:10

everything that it claims that it can on

22:12

the package. But what it's

22:14

claiming on the package is not a tenfold

22:16

increase. It's already only, you know, a 2.5

22:18

fold increase. And

22:22

I don't think that was big enough to move

22:24

the needle enough to make business environments

22:26

go for it. Because

22:29

in business environments, you've got more of a pressure

22:31

to actually test the thing and trial it and

22:33

see if it does what you want it to,

22:36

and nobody's finding that those things are doing what

22:38

they want them to. Well, I think

22:41

also in business, if you're going to upgrade the

22:43

switches and all the infrastructure anyway, only

22:45

going up to 2.5 didn't seem to make sense versus

22:47

going to 10 or 25 or 40 Gigabit. Because

22:51

the other thing is, really to Jim's point, the

22:54

drivers for these new devices are maybe

22:56

not as mature. Like, you can get

22:58

used Mellanox 10 Gigabit stuff that

23:00

was when it was mature, but

23:02

is still five or ten years old now, and

23:05

get the full 10 Gigabit for a card you can get on

23:07

eBay for $15 or $25. And

23:10

my base-net rack is full of those

23:12

Mellanox cards. Like, they connect X3s, they're

23:15

connect X2s, they're super old, but they

23:17

can saturate the whole 10 Gigabits. Whereas,

23:20

yeah, all this 2.5 Gigabit stuff is still

23:22

pretty new, and the drivers aren't as

23:24

mature. And even the hardware is like, well, we're trying to

23:26

make this as cheap as possible, right? At

23:29

the time, the Mellanox is like, no, this is

23:31

a premium product. This is like the best thing.

23:33

We're going to put all these extra horsepower in

23:35

these big, chunky cards. Whereas the

23:37

2.5 Gigabit is like, we want the built-in

23:40

Ethernet on your motherboard that we make as

23:42

cheap as possible to have a bigger number

23:44

so that you'll buy it. The other

23:46

thing I'll warn anybody that's interested in getting into all

23:48

this, it's really cool and exciting to

23:50

get into 10 Gig or faster networking when

23:53

all you've done is Gigabit up until then. And

23:55

Gigabit has been a thing for so long that

23:57

I know for a lot of folks, it's just

23:59

like that. That's the speed that wired networks go.

24:02

And the idea of, oh, I can

24:04

go 10 times faster than that is

24:06

super exciting. And I'm not

24:08

telling anybody not to play with that, but I

24:10

am gonna say, I'm gonna

24:13

share that there are two different types

24:15

of storage admin. There's the one gig

24:17

storage admin who thinks performance doesn't matter

24:19

because the network's always the bottleneck, and

24:21

there's the 10 gig storage admin who

24:23

very quickly discovers, holy shit,

24:25

everything matters now. Because

24:28

it is so easy to get yourself in a

24:30

situation where your network will go faster than one

24:32

gigabit, but your storage won't, and

24:34

as a result, you don't see any

24:37

actual real world difference. So

24:39

be aware that going 10 gig can be a

24:41

little bit of a rabbit hole you end up

24:43

diving down, and you may end up upgrading a

24:45

lot more than you thought you were gonna have

24:47

to before you really see that 10 gig benefit

24:49

the way that you want it to. My

24:51

shock drives suddenly feel very sane. I

24:55

just finished a consulting gig for a

24:57

video post-production house, and

24:59

I'm like, all right, we've got our file server with 100

25:01

gigabit, Nick, it's connected to

25:03

four or five different editing bays with

25:06

a mix of 40 and 100 gigabit,

25:08

and we need to be able to do two

25:10

or three gigabytes per second from each of three

25:12

or four workstations at once from

25:14

this mass. Each

25:17

frame of the video is a separate file that

25:19

are each 30 to 40 megabytes, and

25:22

we need to be able to stream these without

25:25

any stuttering. So we need to

25:27

be able, it's gotta be sustain this many gigabytes

25:29

per second, it doesn't need to be more than

25:31

that, but it needs to be this many, and

25:34

never fall below that many, and

25:36

that's a lot harder to do than

25:38

just installing a bigger Nick. You gotta

25:40

get your storage dialed in, you gotta

25:42

have your compute dialed in, because when

25:44

you're feeding that much network throughput, you

25:47

need to have the PCI Express lanes to

25:49

read from your storage fast enough, as well

25:51

as the storage itself being fast enough. So

25:53

you've got CPU issues, you've got

25:56

storage issues, in addition to

25:58

just throwing a faster Nick at it, doing Ethernet

26:00

like you're going to have to configure jumbo

26:02

frames and that's going to be weird and

26:04

different between different operating systems like I

26:07

just had to mess with all that this weekend for

26:09

a client that did a 10 gig upgrade and didn't

26:11

get the benefit out of it for exactly the reasons

26:13

we're talking about. So I spent

26:15

the whole weekend in there messing around with

26:18

jumbo frames discovering that you know with certain

26:20

configurations their Windows clients would just plain crash

26:23

even though everything was configured the way it

26:25

should be with you know 9000 byte

26:27

jumbo frames all the way around

26:29

like there's so many ways for it to screw

26:32

up like on Windows to get a 9000 packet

26:34

frame you actually have to set it

26:36

at 9015 because Windows also includes the

26:39

15 byte frame header on every frame

26:41

whereas every other operating system doesn't count

26:43

the header when you know they're adding

26:45

together the number of bytes that go

26:48

in so it is a

26:50

giant pain in the butt it can be

26:52

done but just again I'm warning

26:54

everybody out there this is not a

26:56

casual project it can get away from

26:58

you quick. Yeah and especially when you're

27:00

like we were in for a certain

27:02

level of consistency it's like so it

27:04

needs to not just be fast but

27:06

not have any latency spikes and

27:08

we have to have enough headroom that we

27:11

can absorb a spike and not ever have

27:13

the transfer rate fall below the number

27:15

of frames that you'd be able to get per second

27:17

to play it back smoothly to make sure it's coming

27:20

out right and it really is like

27:22

I've done on a server as

27:24

big as two dual ported

27:26

hundred gigabit hundred gigabit cards so

27:29

that was 400 gigabits total pushing

27:31

out of a 128 core machine

27:34

and just proving that it could take video

27:36

files from RAM and check them out the

27:38

network SSL encrypted at 400

27:40

gigabits a second and that

27:42

took a lot of tweaking and tuning and

27:45

the biggest thing when you're getting to any

27:47

of these higher numbers beyond a gigabit is

27:50

you usually need more than one

27:53

flow one TCP connection is

27:55

not going to saturate your 10 or 100 gigabit

27:57

network card because you're going to run out

27:59

of CPU. power. There's one

28:01

flow, you can only use one core. Whereas if

28:03

you make 10 connections, then each of those 10

28:05

connections can execute on a different core at the

28:07

same time and have a different interrupt channel from

28:09

the NIC and you'll be able to do it.

28:12

And so for the video editing thing,

28:14

we had to make each client be using 16 threads

28:18

so that we're pulling this frame and the next

28:20

frame and each in a different thread so that

28:22

we would have enough flows to actually be able

28:24

to deliver that much bandwidth because you couldn't do

28:26

it with a single NFS connection. Right,

28:29

well we'd better get out of here then. Remember

28:31

show at 2.5admins.com if you want to send any

28:34

questions or your feedback. You can

28:36

find me at jrres.com/mastodon. You

28:38

can find me at jrs-s.net/social. And

28:41

I'm at andangie. We'll see you

28:43

next week.