YouTube Transcript:
Bench-marking ZSTD for our Linux distribution!

Skip watching entire videos - get the full transcript, search for keywords, and copy with one click.

AutoDub

Understand YouTube Foreign Videos

Immersive YouTube Voice Translation

Break language barriers, embrace global quality content

Solve Foreign Video Barriers Instantly

Video Transcript

welcome everyone to more open source

linux development stuff here today i

wanted to benchmark set standards the

latest and greatest compression

20 years ago rock nukes was the first to

one of the first to adopt b sip 2 which

at the time 20 years ago was the latest

and greatest but obviously

and now t2 was a one of the first

adapted standard

maybe even the like nearly kind of sort

of first up that's that that's not for

william's new situation um so what that

means is compressing our binary packages

and builds the sources like firefox

compressing those and using sets and not

here i made a video probably three years ago

ago

um in the meantime some other linux

distributions also switched to set

standard for its

extremely outstanding decompression

speed but still maintaining quite some

compression and the

the

so the the thing is i

we have here

like two use cases we have downloading

the sources we are recompressing them

for local storage

because they might come in a lesser

format and also for our mirrors plus

we have the binary packages and

i thought so for the mirror we actually

use like set standard

ultra 20 and so what is confusing i

thought that was the highest maybe i

should have or maybe it's not the highest

highest

because i only now realized the main

page says

ultra by the way ultra is guarded behind

that so the higher levels are protected

and guarded

with ultra to avoid people accidentally

shooting themselves into the foot

because it requires levels 20 plus

maximum 22 require significantly more memory

memory

and this is also why i for probably the

ps3 but in general

vintage and rachel like even 486 right changed

changed

the tower creation

for 32-bit platforms because otherwise

you need too much swap and it obviously

becomes lesser performance it helps

nothing if it's the highest performance

stuff if it needs so much memory that

it's constantly

uh swapping to death

and that is why i changed this just

recently some months ago and also kind

of the point of this video and probably

i should make more presentations like

video stuff for more people tutoring but

anyway today we quickly live streaming

the work because now i've noticed

um so here we use 21 or um

oh i changed this already sorry

we use 16 and 1620 and i was just

toying around here and so i can also

live stream that so it was 20.

so i guess we will not uh

yeah so

the thing is instead of just

ballparkings i think i came to 20 some

three years ago which probably after 20

it became just way too long to compress this

this and

and

16 i don't remember um because

because

i read loss of the point of the videos i

read now in ultra

just i think because we yesterday uh

briefly touched this right and

and

so that sounds like maybe 19 is enough

and also yeah because yesterday i

noticed the power piece so the issue is

now the power pi power pc iso has become

so much larger

and what i actually

quickly aka famous quickly

five hours later

welcome everyone epson

what are you even freaking

posting there anyway

we should actually benchmark that a

little bit more proper theoretically um

it's also a little bit funny that was

all the sets down that fame

they mean they made a

not so old thing i mean basically

probably even i could make a better

video if i wanted to

but it's a little bit funny that this major

major

projects didn't update their stuff

stuff

you would sing such big open source projects

projects

like at facebook theoretically they

could actually update this graph for

each version but i guess they are mostly

a year or two year gcc7 and

i mean i understand compiling this takes

quite some time has seen your life on

i would expect uh people working at

facebook would automate that

so let's take something not as that

that

large or so

maybe let's take lvm

lvm lvm

lvm

that here so is it soc this was epic

thought ripping so this was originally

400 megawatt for eight megabyte

so for

um so that's directly we should create some

some

some nice graph for um

i in

time

and then i prepared to read something

somewhere all the time [Music]

all right it's also like user

yeah it was

capital and what did i use okay

so that what i want to know is um

the point is also test the stuff

professionally right and maybe two yeah

so that so that should be the

a maximum resident set size i

would hope that is all allocations i

mean this is also that do they explain that

that

better yeah so yeah you you really need

to know unix stuff to

be sure anyway

we get all the time so

elapsed real time [Music]

[Music] um

combine that so what we want to know is

how long did it take

so that would be total user time

for multi-processing systems also we

want yeah okay let's let's just use it

like that for now and

maximum memory size so that is um

so this is not even this is not what we

really want to know but let's first

start with that sets.net compress llvm

and writes it to lvm await also with

dash level so that maybe always ultra

i hope that should only unlock this level

level

because otherwise we would need this for

over 20 and

and

so let's see maybe also

yeah controlling the of course [Music]

[Music]

let's print this out

and so theoretically we should make some

nice graph so this is compression the

compression we actually don't care that

much about we just need those files i'm

only i mean my

epic thread ripping

horizon systems you can basically

compress relatively as long as they want

until it becomes too annoying but you

also see already how much the time

increases there

and this might also be why i've chosen

20 and not like 23 or whatever the

maximum was

this is also the memory alright so this

should be the memory

and yeah it's it increasingly

apparently needs more memory i should

actually research and clarify exactly

what the maximum residence set size is

if this is a maximum locations

and so actually i'm more interested in the

the

decompression stage for the various

levels which is also why i'm capturing

those here actually thoroughly so what

we probably should do for proper youtube

video is

like time

uh compressed size

maybe even megabytes per second and

and

time says all right memory

and so we could

better scripting than manually taking

your precious 486

486

and i'm

reinstalling this because when you sit

there a day obviously you don't want to generate

generate

20 isos obviously this is how you test this

this

so i'm mostly interested or i'm only

interested to see if we can use 19

instead of 16 so that our

do not become that large but you see how

long this already takes a few i also

changed recently um

in this massive

optimization and improvement phase of t2 here

here

three years live on youtube i also

changed our build scripts to compressors

in the background because obviously if

you compress such large um

such large archives uh you know what

maybe we could even

or maybe not um yeah because because

obviously you stall you build quite a

lot right you can already

cook although this should be relatively uh

parallelizable with this mighty thread

implementation but

it's still um the next configure can

even even if your epic thread ripping 32

thread machine is

saturated with a stationary compression

which actually by the way we

we could await a second voice it's not

uh darn we didn't okay this was stupid

you know what

and also press do some

buttons we didn't actually we don't even

multi-thread that so we actually could

have split this up

but yeah there's also i was already

wondering whether why does that slow but

yeah you see 15 to 16 it's like nearly

doubling the time already right or 14 to

60. let's actually control see that let's

let's

do that fully mighty threaded

could it be

i realize this so t

should not be t

or is only compression

compression [Music]

the [Music]

compression level does it not

so yeah that should be obviously embarrassingly

it's surprisingly not fully texting the

interesting

even more of a point to continue [Music]

[Music]

course instead of threads

or is it simply it could also be that

this is simply not more further

threadable

yeah thanks to graver i i figured as

much things um so yeah you see how much but

but

it's crazy i mean this is the stuff you

always think like yeah it's like

fully mighty threaded and then only live

on youtube like hold on a minute it's

not like as perfect as i

hoped it would be anyway

anyway

let's test so memory wise here it's

going further so this is only

using four max of memory surprisingly if

oh wait a second or is it in okay maybe

we should actually check the main page

in kilobytes okay so [Music]

[Music]

this is okay that is actually then quite

some significant more memory

incented

what was it i said four megabytes and it is

uh four gigabytes

is it 4.7 freaking gigabyte

and welcome everyone um

yeah so they also see why i didn't use

look how slow that is right 21 is still

apparently voice is not further

not really for the mighty threading interesting

interesting

yeah that also explains why i did

because last night i was wondering why

didn't i use more but no i think i

vaguely remember like it became super

slow is that what our benchmark

benchmarks are for right um

so this was 21

still do with what was even

what's this one gigabyte i'm slightly

if if that is true how we calculate

calculate

the gigabyte megabytes from that

[Music]

because the archive was 400 megabyte or

it should be significantly faster than

three years ago when i tested this it

was always in the news like new

optimization more optimization

better compression more optimization so

but of course sometimes also people

should actually

how large is it even oh wait a second

can we finish that

what you're saying with dash m for

uh memory

the memory usage limit by default does

not use the 128 max of full decompression

decompression [Music]

at least worst case with some

swap space

i wouldn't

really further limit this we obviously won't

won't

quote some significant compression

compression

um but thanks for that pointer i mean

one thing that would be cool is we could

theoretically like if we wanted to

optimize the heck extremely out of it we

could theoretically

let it create a dictionary from all

eyes or so

but that is a little bit too much so

basically we could have it like this

training set for this dictionary stuff and

and

like like pre-train it um

then probably dictionary builder year um

i mean very theoretically we what we

actually could do

but i mean it's getting a little bit

complex then right i mean maybe i mean

it would certainly be an interesting

youtube video potentially how much we

could get the dawn i mean not to delay um

not to delay building

building

all the packages at the end we could

actually train it with the first binary

package we built

like say the first package might be zero

zero d3 like this file system hierarchy

standard so it doesn't have a binary so

that wouldn't use a dictionary

but all the other

packages like like the first thing

installing some binary like julipsi or

stuff have it trained that that

certainly should be quite like the

dictionary is matching for arm

executables versus versus x86

um dictionaries and

and

so then have once the first dictionary

is built

from the first binary package installing

something binary you need to have all

the other stuff built with that

dictionary and then included into this

either that would be

potentially some

interesting uh endeavor for for rainy uh

it's not my priority i only wanted

okay so actually

23 doesn't exist

fair enough so

also maybe we should have how much

so we have two

two

there's even the time here

so that is the size so the memory use actually

actually

interesting the memory was was even

slightly lower by some

pure chance and stuff but

how long did it take 170 yes so it took

even longer so basically we are right

we are right now using 20.

interesting that

was 20 i mean i thought you have some

control characters

so i

it's crazy that's a slightly faster that

maybe i mean it could always be that

due to a larger window

window

it's compressing more far more earlier

or something

it sometimes matches more and compresses more

so this is a crazy find really to be honest

honest

yeah so

is it significantly smaller

yeah so basically we should actually for

proper youtube channel

this is basically just

just um

oh this is how would be and i thought it

was by size

so yes the size is

pretty sorted as you would expect and i

mean to be fair it

squeezes quite some additional

additional

compression out of there to be honest i

mean this is actually

the i mean it's actually kind of

impressive that

i mean it's three let's make about three

megabyte here and

i mean 22

it is actually crazy better compression

i mean 10 freaking megabytes i give them

and this is actually freaking impressive

to be honest

and this is the stuff that i love right

not this like ibm breaking bin utils for

page sizes and other random london's i

mean the set synapse is really

really amazing compression technology i

did not expect that

kind of advancements

um that is really cool

should we actually yeah so let's test

decompression and then

make a decision so right now we are

using 16 i mean we of course see here so

right now our 16

yeah so that is 15

13 megabytes larger than

i'm kind of thinking we could even

increase that to at least 20. i mean i

mean this level is really impressive

let's see what the decompression um

memory usage is and theoretically

there's also some win longer window mode

i wonder if

maybe we should do the same i mean it probably

probably

yeah let's let's just run this for like

some experiment when we're done with

this stuff so right now i only want to

determine the

decompression memory use and

so we don't need to save that so that

decompressive stuff there

also best with

like that so let's see

okay so this um

okay 23 written i mean this

decompression is so blazing fast it is

actually crazy

um i think it doesn't support the com

thread decompression does it let's

fall in of all of that stuff

that was t0

but i think it's not multi-threaded it's

decompression yet i wonder maybe it's

not even easy um

to multi-thread so yeah um that is

extremely fast as you see so if this are

kilobytes that would be 41 megabyte

so that is 19. yes so i mean these and

the difference between actually here 16

is actually significantly lower

i mean

so that just wait a second so that is 25

but we are still speaking if this are kilobyte

kilobyte

wait a second i'm getting confused with

um [Music]

[Music]

so they said kilobyte

uh getting already

wait this should be kilobytes so that is parts

parts 25

25 megabytes um

i mean this is actually not unreasonable um

for a large package if that is what it is

do that again with

it if not

not

i'm confused but it even has an effect

this is a little bit yeah people always

say read the

effing menu all right and then

i wonder

i read it and wonder

actually did it because they're not equal

equal

it's a little bit

let's maybe should we

it's actually

how much

better as it apparently is i mean we

will see with the memory requirements i

i mean this needed

that's quite far away yeah i mean i wonder

wonder

but the thing is i mean doing that once

it's already quite some time as you see

your life and making nice bar grows even

takes more time we are twenty

gear so that is that is 10 freaking mega

alexander says training as a dictionary

using all available archives i wouldn't

use all i mean um

there's also a limit i'm i know from

machine learning i started also quite

some machine learning

getting a little bit dark you don't see

this on camera the camera is so bright

actually it's pretty dim here it's uh

my eyes start to hurt from from the

contrast of the dark background

it's a little bit overcast here um

so you can't you can't train a

dictionary i think it's probably enough

with all archives um this is too much

training you can't like put the

knowledge and stuff you could also

directly train that then i think it's

enough for this dictory stuff to

somewhat recognize common patterns

i mean of course it will not like

intelligently recognize patterns like

power pc versus risk five assembly

but it will do so by dictionary

statistically statistical heuristics so it

should be enough

with with

one of the first like binaries

any binary should be enough to recognize

common like the f header um also the

various sections some strings some

common all the

jumps and moves and stuff all the bit

patterns that all those

most current occurring stuff um

leave there in the binary

especially you think rms previous video

right i had says this arm instructions architecture

architecture

this old fashioned 32-bit arm so many

unused bits that are often

yeah let's see how the deep compression

goes i yeah i

wonder if that will i'm actually

surprised how

how small the memory usage was for

for decompression

i mean obviously the compression needs

to build all the dictionary in memory

and certainly where some

dual channel or quad and stuff memory

and it comes in and certainly plenty of that

that

here is not it changed significantly

though look at that i think previously

it went up significantly after 16

and now um

it actually significantly went up well

at 14 15 already decreased

decreased

probably due to better earlier more

22. yeah

i want to run the next thing

okay let's take a look at the sizes here

while we do that

are you freaking kidding me do you want

it oh wait no this is in in progress

right this can't be other yeah this is progress

progress

um 21 so yeah 10 10 freaking megabytes

smaller right

that is crazy

i'm kind of tempted to

last minute even rebuild all those isis

again i mean it's fully automated right

i just leave the epic threat ripping

horizon here running for

uh for 30 hours and i have all the isos built

built

as we cross compile everything

so anything else we wanted to

i did not

[Music]

there's any reason that it doesn't download

download

ah come on we need

we we are ready for

and that is the thing right do i buy a

threat ripper i need also single

threaded performance right and

i think the threadripper will not really

give me more single threaded performance

so probably for content making a proper

review video

i don't think only it

and where are we wait a second released

oh did it again i did this with okay

okay um

so yeah i'm so ready for i did not wait what

yeah

so ready for the next sun architecture

so um

so memory wise

yeah i mean this

was it

seriously running for wait a second this

is accumulated so this is not wall clock

time this is accumulated multi-threading

multi-threading

i mean yeah i mean it's excessive um

otherwise i mean we

care more about decompression let's

let's do them

decompress

let's see what happens

yeah it needs more memory though i think

yeah i think this significantly

previously had 25 max of memory right so

it's no wonder that

here is a decompression right one and

two decompress

yeah so yeah we probably can't do that

so that was i mean we are not using 11

megabyte but

so we previously had 16 so that would be 25.

25.

i would even say 19 is 41 should be fine

so i guess we can use the regular

19 for 32-bit architectures

i mean you can even decompress that if

you have a high-end 486

if if there is something like a

high-end 486. um

i guess it's fine even on the p3

so we we have some my memory free after

decompression in it already

and you can actually wait swap so yeah

so we can do that i would say and yes so

we can't do that though um

we can't do long however

i i think what we will do

is for 64-bit mode because this long leaf

leaf

is so so how much it was at 500

something right

even the octane i mean yeah there is a

thing i mean that is quite

but i would say all reasonable 64-bit platforms

i mean why not make it more convenient

for everyone right

i like to support vintage retro x cases

edge cases but in

in a reasonable way i would say

and i would say

512 ish max of memory

means the only thing is sga octane

but maybe we are using certitude anyway

because we default as you obtain really

i mean yeah okay i mean even

because even ultra spark i mean i mean

yeah for vintage if you i mean the thing

is supporting 20 year old

old 64 ultra spark stuff even i didn't

have so much memory in my ultra spark

maybe it's not

it's a very

let's think what else do we have we have

okay pa risk we built 32 but user land um

what else do we have alpha titanium

titanium

i mean it's no wonder that with with

such memory usage and like 512

megabyte of long window that

we get so much better compression

basically than one long is as good as four

four five

um okay wait so that ah this is why the

32 didn't delete because

it was 23.

how much i mean 20 too long actually it doesn't

doesn't

i mean given that okay all those are so

basically 18

i mean what this also means is for you

you can significantly improve

compression without investing that many

cpu cycles as 18

long as here in this case equivalent to

so previously we had so 22 okay i mean

what was i think 22 that

yeah so we use this is decompressing

okay this is

this was already using significant

memory in the 22 case

so this is why this doesn't improve

further because it was already using

using

so yeah what does it mean for us it

means for us this was 21

that was 270 what what have we been

we were using 20

and 21

gives us a little bit 21 long i mean 21 long

sorry is basically

what was it zero point i mean the

decompression time is kinda constant

actually this long decompression time is

at times

in the lower level actually faster

i mean if you use so much memory and

uh default is 27 or so window

window

defaults does this increase the window

size whenever memory is supposed to

compress and decompress it

proof compression right from doing something

something um

i wonder what results we get when we set

for like 10 to 22 compress the stuff with

with

this window must already have some value

that is [Music]

[Music]

so 10 yeah so this is also verse okay i

think it doesn't make sense to manually

specify another

okay what does it mean for us we don't wanna

wanna

we want to be able to install in

that means our

20 choice wasn't better

i mean yeah 20. i mean there was a

maybe i should yeah anyway probably we

we kind of leave it like that i guess

because 20 ultra but good it's good that

so that means for us we can safely use

19. so it's still an improvement and

that's also often a thing right you

think here we need to double check that

and prove something and then it doesn't

improve but we at least we understand

this is not better um

also something you wouldn't get

immediately from the manual page

especially this window you would think

that's always making it longer but

apparently sometimes it's if you specify

it's already shorter than the

the only thing is we

i mean for the building it doesn't

change our mirror default mirror compression

i mean it would be significantly smaller

right 10 max of less

maybe we could increase that a little bit

i mean if you build i mean

unless you build on vintage systems again

again

okay you know what let's play it a

little bit safe

we did some research with 20 is

yeah i mean that you can still use t2 to

download to download and build at least

reasonable small medium-sized stuff on

vintage hardware here i guess

some compromise

now that we did some research with the

memory usage

i have significantly smaller isos again

yep that's probably for this finding um

i hope you found this interesting probably

in the future i need to make proper

graphs maybe even live i know i need

some better graphing tools obviously

if you have

tips up in the 20 years ago i was using

plot equals but someone's also probably

not really maintained anymore you can

actually check that also mean it's not

the most outstanding it was just

better for what i use than a lot of stuff

but i think as the last i checked

probably the last update is decades ago 242

242 yeah

yeah

news for more you click news on like a

okay so loss of sounds they have up and

down this for some new poison stuff or something

something um

what's the greatest poison fan

good for them um

not good for me but it's there maybe

should we package that

let's put this on the to-do can i have

[Music]

the future of the looks desktop the copy

anyway but that's it then for today i

i mean i could recompress this [Music]

[Music]

maybe i just manually recompress those

packages maybe i'm not feeling like

rebuilding everything just for a second but

but anyway

anyway

that's it for this video let's

and

should we do that

nice so how much would be safe you have

925 megabyte for

f in time

time

f removed

why is that

remove thing

yeah

all right so it could be consistent with

projects like you as it's also a thing

right that i tried a little bit and it's

encourage people to work a little bit more

more automated

automated versioned

versioned

reading the manual or i think stuff

through and not always yellowing

everything together

and um

a little critical and long term

and surrounded sinking there it's very

very it's not the best practice to

recompress the stuff but

but

for one time thing

i think it's okay i'll also probably

only do this for the power pci so not

for the other ones

that was some of the largest

maybe the others are not firefox and stuff

is it even fully mighty yeah those

mighty threading is

i love my wish you could see my rgb

animations on

my cpu utilizations memory sticks

year was actually thinking we could

recompress this more in the background

actually run all of them in parallel

maybe i should have done this

now i forgot already how much was it 100

yep but for the most part so it's not is amazing

i really wonder yeah we probably maybe

we make this another video trying this i

actually didn't try this dictionary

based training thing there

and that's then we probably better do

where are we linux i mean there are

so very soon the further updates data is done

done

data certainly also the best invention

yeah but so this workload also doesn't

uh profit from uh threadripper right so

the horizon here is as you see this

this

parallel compression stuff only goes so far

far

i'm actually surprised it doesn't do more

maybe we should create a freaking manual

compressing using threads

detecting a physical cpu cores

[Music] hmm

okay now we're done

yeah so from

125 so that is over 57 so 75 megabyte

that is certainly something

the only other thing i wonder is how

much the mighty threading is reducing

the compression but

but

i guess we tested enough for today um

so yeah

that is the summary and how to widen the

pros and cons of this stuff ultra 20

and i'm kinder in the flow i wonder if

test how much

i don't i don't really have the time now

use

set standard nineteen [Applause]

how much was it this was

800 was it 850 of 925 i think [Applause]

[Applause]

i said standard 19 instead of 16

16 f4

okay [Applause]

something of that i would say um

um so yeah that's it for those videos the

so yeah that's it for those videos the only other things it's only with the

only other things it's only with the dictionary stuff

dictionary stuff we do this another video maybe i should

we do this another video maybe i should actually maybe i make a note otherwise i

actually maybe i make a note otherwise i might eventually forget about this

there are additional complications not only we would need some surrounding

only we would need some surrounding scripting for training the dictionary

scripting for training the dictionary and then making sure obviously to use

and then making sure obviously to use the same dictionary for all tables and

the same dictionary for all tables and additionally changing our iso creation

additionally changing our iso creation to include actually this dictionary

to include actually this dictionary right my understanding is

right my understanding is um you need this dictionary then um

um you need this dictionary then um with

with in your iso

in your iso so

so um it would need to require a little bit

um it would need to require a little bit of surround scripting and tooling to

of surround scripting and tooling to make sure this and even the installer

make sure this and even the installer right we would even need to patch we

right we would even need to patch we would even need to patch our freaking

would even need to patch our freaking outdated vintage mine binary turbo

outdated vintage mine binary turbo extraction thing to

extraction thing to run

run runs that standard with

runs that standard with just this dictionary so

just this dictionary so yeah probably not really worth it we

yeah probably not really worth it we will test it another day obviously

will test it another day obviously because just to test it and

because just to test it and be smarter and learn something

be smarter and learn something hope you don't forget to subscribe

hope you don't forget to subscribe have a great day or night and i hope to

have a great day or night and i hope to see you soon for the next fun stuff to

see you soon for the next fun stuff to come

Click on any text or timestamp to jump to that moment in the video

Most transcripts ready in under 5 seconds

One-Click Copy125+ LanguagesSearch ContentJump to Timestamps

Paste YouTube URL

Enter any YouTube video link to get the full transcript

Most transcripts ready in under 5 seconds

Get Our Chrome Extension

Get transcripts instantly without leaving YouTube. Install our Chrome extension for one-click access to any video's transcript directly on the watch page.

Add to Chrome — Free

Works with YouTube, Coursera, Udemy and more educational platforms

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube

←

→

↻

https://www.youtube.com/watch?v=UF8uR6Z6KLc

YoutubeToText

←

→

↻

https://youtubetotext.net/watch?v=UF8uR6Z6KLc

YouTube TranscriptPreparing your results…

YouTube Transcript:Bench-marking ZSTD for our Linux distribution!

AutoDub

Video Transcript

Paste YouTube URL

Transcript Extraction Form

Get Our Chrome Extension

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube Transcript:
Bench-marking ZSTD for our Linux distribution!