Rust Programming Language Rust Malaysia, 10 Aug 2023

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: Rust Lang Meetup - August 2023

Description

Zheng Yan (zypeh) will be sharing about io uring.

A

Hey guys, I think we're going to start.

A

Talking down, okay, so first of all, I want to thank each other one of you uh for attending this, so really appreciate you guys take your time out from the busy schedule, especially after working our. So allow me to use myself. My name is Jeffrey I.

B

Want to organize office Malaysia.

A

Together with Ivan over here, so what happened? I've been doing this for quite a few years, so yeah we have to meet up like every two months um yep. um So today, uh our main agenda. We will have a talk uh by uh about Iowa urine right, um so again appreciate for you to really give a talk and effort and preparations um so before that before we get into the topic, I would like to spend some time in our last Corner talking about some decent update uh in in Iraq.

A

So, okay, uh this is our social media, so you can really find us in the telegram Channel I. Think most of you are in here. Joining is a go like Channel, but we do post some updates as well and we do have a rock Malaysia Facebook group, Russia Malaysia Facebook group- uh it's not very active, but we started that we still post some upcoming updates or events. uh So if you guys I want to update our kids, our children can join our wrestling, Malaysia and Okay.

A

So so, first of all, let's start with our grass 1.07 uh versions, uh which is being released I think last month right. So there are some key updates in this version. One of them are the first first one is a spouse bike.

A

So uh if you guys knows right so basically in short,.

A

You will first need to verify the version and dependency that you use in your program. So what how it does is you actually download the index of all upgrades, which is all the packages right so on all these index actually hosted in GitHub repository, which is in the old way. So every time when you try to run a rust program, you're actually sort of like get cloned down the whole copy of indexes in your local machine. So with this actually causes.

A

Some issue is when the index grow right and because everything you've got people, publish or update their positions right and when you download or update your local, create index. It will take longer time.

A

So when you try to deploy in your cicd you'll talk about uh amount of time so with spots, I really change the it's sort of like an indexing protocol, so the way that they actually enhance this process is like they, instead of using git flow, no downloading indexes they're using https, and all these uh sparse in-depths are actually sold in a power system which you can access through index.creates, IO yeah, so so previously in 1.6 something version. uh If you want to use fast, you have to explicitly config right now in this latest version.

A

It comes by default, okay, so basically, this is a box bar, and then we have our one cell one slot. uh It's basically uh started with a complete Community Driven packages. uh uh Basically, in short, allows you to create substance like Global variable in Rust right, so um once locked, it's uh it's a counterpart of one cell. One cell is mostly threaded threat, safe and one cell is for single thread right, and so in this version they actually they're stabilizable this great and they added to the standard Library.

A

So you do not need to install uh use it. uh One thing other packages: I can use it straight from the standard library, and then we have some method called instrumental.

A

Basically, just to differentiate whether you are in running screen or you are in interactive mode right. So it's helpful if you are working on two weeks right so yeah, and then we have another version of class 1.71. So uh some key updates here is the debuggle visualization attribute. So uh this is sort of like a new attribute that you can add, switch support, uh Navy's fault. So when this focus is something like a screen used by Visual Studio code from an evaporate inspection output right so right now they support it.

A

So this files actually contain a format of what kind of output. When the user inspects the data structures, what is you can customize the format so right now you can add this attribute, and then we have a const initialized trade, local. Basically, you can use cons, block to initialize uh variables inside three locals, so but this uh this actually, this feature actually has been a while since in rocks it's not that they did add it into documentation. So a lot of people are not aware of it.

A

So in this latest version they add into documentation stuff, and then we have battery 0.11. If you guys don't know battery is uh open source uh Ras uh game engine right, so it's written fully in rough and they have a lot of new updates on this Lead You by Walmart versions are like performance, Improvement right. We don't like that right. So, um if you guys are interested in various game, development feel free to go and have a look um so collapse. Yeah.

A

Some exciting news in the rust Community right uh more job opportunity, if you guys want to be a rust developer. So we have new open source data as an open source company right just three years old, open source company, and they recently raised four million around 4 million of fund and they started the company so in short, polars. It's sort of like an alternative to tighten data frame. So if anyone are in this I'm sorry pandas, yeah pandas, so it alternatively pandas in the python yeah.

A

You might use uh pandas right frequently, especially for data scientists right, so these are sort of an alternative to that, but it's written in Rust. So we had inherit all these performance from Russ and it's using everything, Apache arrow in memory, data format, all those kind of stuff. So yeah, it's pretty cool. You guys can try it out. It's really amazing yeah. So for for us, what's the difference.

C

Watching from YouTube might be, that might be interested.

C

As in products have the ability to have the lazy data frame because foreign.

C

So what difference is that the lazy, the evaluation as in the plus you can you can load partly in memory, so it doesn't load everything into memory. It can go through your go through your data easily like go through something like that, and the benefit for us is that for us is able to use multiple threats, because it's designed from scratch to do like the outdoor press and make it faster motivation about the process, I'm, not sure, but it can run properly.

C

That's why it's faster and then more like less memory use, but compared to the integration of, and that's maybe it's lacking a bit right now. Actually it doesn't, it doesn't really integrate, but pure.

C

D

C

Eagerly the data frame and the lazy loaded you just like one line so just one line that you can just make it lazy and then one line that you can like performance, like oh jump,.

A

Yeah, it's still very young like three three only three years: um okay, uh the next one is that cash, uh it's something in the crypto World 3 right, so that catch is basically my understanding of it is like a point, but with privacy implement the transaction actually public right. So some of the scientists and storage is an issue. They fought Bitcoin early and they added its privacy features uh which actually uh encrypt all your personal financial information.

A

So basically itself is a cryptocurrency and they have their own network, so um they they recently I think they released their zebra, which is like they are note the note in their Network yeah, it's written in Russia, holding so yeah, so you guys are interested in cryptocurrency number. Three.

B

A

In terms of how we Implement writing a uh no, you guys can have a look.

A

um Okay, so the next one is a diesel 2.1.0, so diesel, uh it's a Corey builder orm in Russia right so who say: Russ is not suitable for web development right or backend right. Basically, you have everything right. You have some web framework. You have Diesel with all this kind of stuff and even have a dioxidants right for front end. So we can basically do the full stack using rust right. So what interesting in diesel 2.1.0? Is they release a new CLI uh features?

A

uh Migration features where you can migrate your screen by create a migration generator accessory based on the rasa logo like we Define the local restaurant, and then you run this deep screen. It will check the difference between your database schema and your local schema, and then it will generate the SQL schema for you. Then you can run and migrate. Yeah database, okay,.

C

But the concept is more like towards the back: it's not like the active record pattern, but more like towards the data. One is called yeah and it's similar to heat structures in a way that it is. The code looks similar to how you write our sgr, except the API and.

C

One thing I like about that is that, because you're right as you write the code, then the compiler will compile the code for you. It's like all. You write that all a lot of functions like this table not insert not something, and then the compiler, the compound right when you cover the window or it works and he's just using the normal normal function. Although we have some code generated in the back.

C

Sqlx is a bit more towards writing. Raw 0.. Let's go X One Thing interesting about SQL X. Is that uh because.

D

C

People might not want to write poem because it's tied down to the platform, although I find this quite interesting as it solves a lot of issues with other words. Otherwise, sometimes you write age cases code that is not in the documentation.

C

It doesn't work, but this is everything that encoded in the type. So the query drawing that join those then, but there's downside to that documentary are very confusing the types that many many paintings are together. That's the downside, but then they do it the type safe not just for postgres. They do it for multiple databases, so they're just doing it for possible. Syntax then deserve you support the type for that, like let's say you can use, upset I, think you can use after one or one or two of their base, one some other devices.

C

You can use option and Diesel put that in the type as well. So in type when you, when you use a SQL database, you have to answer so I find that quite interesting and the SQL X is the alternative to discern that. You write across SQL the difference. The sequence and interesting idea is that SQL X, you test the code against your database during common time. So when you compile you test whether the query work or not, foreign.

E

uh Before executed,.

E

Let's say you, you update any.

E

About testing on its own.

E

C

N't running a person but I I think I'll. Try it outside.

C

A

Christmas, oh then, the last one is uh do some some of this in the community. uh There are some proposals with upgrades. Io policies are so uh this actually caused a little bit of Dramas in the rust Community itself, um um so they actually tried to tackle. uh One of the problems we found in the crates itself seems wanting to actually reserve the name. They published a packet reserve, the name because we don't have any space in the last phrase right.

A

So if you use the name right, it's first come first matter you can't people use this name. If you like, XXX and people already publish package access actually cannot use it. So there are some people like purpose, doing this to reserve name, and then you need to do a lot of stuff like technology transfer right. So basically they come up with a new policy to tackle this problem, saying that if they found you are not uh actively developed right on your, your publisher packages creates right.

A

They will consider you against the policy that maybe they can delete your character, so this uh I think I think we solve these issues right like how ghost does it right? You can have my ninja affiliate, slash, Tokyo right, then you have everyone can cast their own Turkey right just using in space I'm, not sure why I think there are proposal of new.

F

Spacing as well so, let's see how it goes up. So if.

A

That's interesting, this topic I go with that GitHub uh pull requests.

A

Basically, like, how do you define like okay with your activities.

A

Right so yeah so yeah, so basically that's it. um Foreign.

F

uh I should share my screen. I actually share my screen here with this cohort.

F

Let me share my screen.

E

B

F

Screen should be customers.

E

F

F

So is anyone familiar with the IOU ring or even heard of the name.

F

F

To Mexico, okay, no problem, so my talk.

E

F

Yeah, okay, so my name is.

F

My initial name, so what I do is I am a software engineer.

F

A software engineer working in an automated trading firm. So me personally, is coming from a functional programming ways and functional programming.

F

You keep please yes, so previously. I am writing Haskell and all this. So when we are writing Haskell, we are conscious about the runtime of this, but because of the nature of the language itself, it handles all of the um IO and how the language um with the real world like IO. So that is why the name called IO and the urine that they will explain to it. The IOU ring is a Linux specific, a new mechanism, a new mechanism to help you to do all this at home in a synchronous way.

F

So whenever you want to send a packet in a network, you want to write to a file. You want to read a file anything that the machines need time to actually does it to process it, even if you are moving some memory from one end to another end of so.

F

I'm, the one who will be presenting the IOU ring and explain to you in that on how iodine works and what is the new case of it. So me I am an electronic trading and enjoyer. Why? Because it Packers a lot of problems where, when I'm working in a startup doing all those ciud stuff and points- and all this and I find that doing the electronic trading and other military will actually gives me a lot of experience and a lot of problem.

F

So what I do in the my company is I took HPC stuff, which PC stands for high performance computing, so we deal with low latency manner of code because the market is so fast, so we have to head fast to actually process this and then we're not necessarily to be fast. We need to be a low emergency because Market sometimes can be inductive. Sometimes it can be very active. So we need to process these packets as fast as responsive k.

F

So most of the time when you heard about electron trading is, should be Buy. Low sell, High I mean how do we earn money? Yes, we do Buy, Low, sell High, that's how we earn money and then one of the way we want to earn money is actually doing Arbitrage.

F

You act faster than people to buy things in a cheaper way and then you add faster than people to sell it in a higher price. So that's how we do but a big chunk of our life as a electronic Trader. We are doing Market making. So that's why I need to explain to you what is market making.

F

So let's say if you love The Ring fan. If someone wants to buy something in real life the world, not in the middle of things in the river, if you try to buy something and then if there's no seller, what we will do you can't? You can't do anything because there's no salary to sell things, so that is, someone needs to responsible for selling. You things the bigger to make a seller of this thing. So in order to have this kind of Trades happen, you have to match these two orders by itself things.

F

So the market maker comes up is, is the one who doing like a Supermarket? You hold the inventory for you, and then you sell things yeah, so the risk they got is that you're going to release because you buy something and then you need to sell it. You'll hold it. You need to hold it and what, if the price wrong, and then you have to sell, it means that you buy high and you sell low, that's the thing.

F

So what we do is we provide the liquidity because someone if they want to sell things they don't have a buyer, then Hub. We provided liquidity in the financial Market and then second thing well in the in software engineer perfectly we are is our requirement, so we are highly sensitive to the price and also price of the market. So we need to keep eyes on it. We need to listen to a web, socket away from the exchange. There will be a market free coming to us toward us.

F

So that's what we do for the price update if you want to do the cross exchange that got at least three okay and then for those assets. If you want to process multiple assets, you must have multiple websocket of it from different exchange from different endpoint, maybe from different Center and all this thing. So we have to listen to for every trade. We have to listen to all the changes stream for the book stream and then once you execute it got the update to us 12th the executive, Stream So.

F

The crypto exchanges will make us make a lot websocket connections, so this kind of connections. um It may be easy when you are doing it known in an asynchronous way. But if you are conscious about the latency, it may be a problem because, let's say you have hundreds of web software connections and then you will keep pulling the sockets to see. If there is a update or not in the system way in the OS, you need to keep pulling it keep pulling it and all this so for us.

F

So for us it runs. There is one Library called Tokyo for Tokyo. It is the defective implementation in Rust. So whenever, when you are in Russ, people will talk about a.

B

F

There's only one mainstream implementation, which is occur, it almost monopolies, the whole ecosystem and then the ball Plus.

F

Yes, we do have a Syntax for the asynchronous from a sequence programming, that's the same and awake just like the JavaScript. So whenever you so people who write JavaScript will find the runs. It's actually quite interesting because we do have a thing and a way and for function. That requires you to avoid things. You have to um hiring it. You need to prefix it as C, so this is a sync function inside it can use the way.

F

So what does a simple way does? Is you insert a way you purposely help the OS on your tablet tells the program to wait for it to happen, for this is a synchronous block.

F

Whenever you use the awaiting the process won't be processed until there is a web socket connections packet coming in. That's what you do. Let us sing the receive. The receive method is actually a c function.

F

That's how you can read it so we're going to read it and then, if there is some message, then you process a message. So why will people will use the talk, show just like what you're doing in JavaScript. You want the task to be um the ones that asynchronous functions to be working as sequence. If avoid things show, the next thing wouldn't be happening before this one finish.

F

So this is how you do.

F

Wait: yeah yeah.

E

F

F

Yeah Okay, the third Point here. The goal here is to eliminate the OS trade. That's one thing that works similar to this one, which is the OS trade. So you spawn the trade. It can actually really work the background, but it sure the Dockers actually uses things. Why is that? Because if you receive the thing, anyone knows how Tokyo works.

F

All right, uh I think I'll just explained the always interested in the documenting. So if I am the one who write the OS trade, imagine a world without a talk show and without the same so we will spawn always trade and then he will do the receiving and then once the receive actually finish.

F

How do you know we finish? We keep pulling it. That is one system called in C language. You need to tell the OS to pull the data from this file descriptor, because you establish a web song uh web connection. You get a form descriptor whenever.

D

F

File, you get a file descriptor. This is a file in OS that uh in other days you this is the handle of this task. So you can deal with this tasks. This FD is a unsigned here. Number you deal with this thing: bring it in Tokyo illness not necessarily needs to pull it.

F

Why Tokyo do have two components? One is Executor, one is a weaker quicker will cool for you. He actually will listen to the OS, listen for the event, SD wages, another fourth descriptor from the OS is not necessarily has to pull it, because the OS will tell you this notify you and then the weaker will break up this doctor us this occurred personally will happen, telling you this away and then you can process it.

F

So what we do when we go multiple.

F

Multiple web functions because, let's say we listen to Two web software connections and then the frequency is different. One is one unit and then another one is two minutes. Then you were blocking until the one minute uh even is blocking because you're wrong, because it's blocking so you need blocking and I will wait for it to happen and then, until the now that the next line of this low need to wait, another connections happen. So what we will do is we have a doctor's spot to respond this a sync task.

F

A single move tells us that this is a computation that will be moved to a doctoral traits with Dr spawn one, and then that is the combination we do so in the line, one which is the second block. The line one is actually making a Channel as a TX partition, RS receiver, so you receive things from the DX, so the TX will be.

F

The landform is TX. The TX send the message and then from the nine that's is rx will be receiving the the message so in the loop away from the mpsc channel to happens so that this second to sixth line will be installed. In the background is background moving and then it's not blocking immediately. It will print this photo this line out and then either look blocking, especially the infinite Loop, and then it keeps pulling down.

F

Yes, yes, a actor is one of the Paradigm to actually deal with asynchronous, but yes, this is something that actor take your example. Yes,.

G

But this moment.

F

For an actor, it actually needs to pass the message right, you pass, the magician here is actually one one uh one. They call bridge to electric transfer the data out. Yes, yeah see me here.

F

So we are dealing with the doctoral runtime and then there are some of the problems where, when we write like this, there will be some problem in when we are dealing with the low latency. But if you are doing with the web server link, this um shouldn't be a problem.

F

Okay, that's one thing: we let the talk show to actually deal with the time. How do they pull the data? How do they process the data? How do they? What if there are a lot of doctoral traits- and there is only one CPU- will does the job, so the doctor will actually help you to manage all the CPU time slots.

F

They will have a queue of the for this task and then, when you submit a task that will maintain a queue, the queue is all those the tasks we sub slot. So whenever there's a result, you're right to be slow and then tells the Tokyo okay, this is what happens. Okay, what if, in the same time there are multiple slots, get written in it? The procure will randomly choose one and then pass it to you.

F

So in the loop you try to receive things, it wouldn't be um so precise, as in you get the result, it wouldn't be so business because Tokyo will handle it for you and it depends on the doctoral run times charitable state. Okay. So this, why number one is a problem, because we let the Tokyo runtime to deal with the pulling kind of times and then to let the doctor to do the cancellations of tasks and then because Tokyo will do it for us. We want to bypass it.

F

We will use to try to see which is another not blocking function. So this wouldn't be a risk. This receipt wouldn't be a sync function. Already we use a try receive so what the try resistance is. It will pull the result. Nco is empty, so empty. What we do we skip it. So in a BC Loop we try to say things, keep pulling it, keep pulling it and then check the result.

F

That's what we do, but this will have a problem because we bypass the vehicle. The Wicker is the one who listens to the kernels of descriptor, the event happy so now bye. We need to pull, we will occur, a system called chord, pulling keep pulling the kernel and tell us what is the result of it.

F

So this is five percent and then I got a parenthesis. Yeah. Beware of this this course back, because you keep spamming. The color second thing is you change the mpsc channel to the third party implementation, because why.

F

From the first line, the docker scene, mpsc amount is a part of the Tokyo runtime is implementation. It used the docules weaker to do a signaling to Signal the another end of the process that oh, that I already wrote, One data, so you need to fetch it away the web called notification and then keep putting it. So you want to eliminate this one so that our 8 to 12 line this low will be busy speed instead of a weight of the result.

F

So there are a few data structure. This data structure. There are a few data structures that does this job without using the single weight, which is is true that man and a Cross Beam, and then um we uh me me in my company- will test these two implementation already and then they are a few of the peak force in there when you are dealing with a low latency code, which is the first one. The test map test map is actually a concurrent test map.

F

So he, when you are dealing with multiple trades, it will include the church starting. You will shut all these um part of the hatchback into different tracks.

F

So when you hit the catch miss you need to check from another trace. It will include latency, which is unexpected, so we try to avoid it for the Cross Beam it got the tail latency.

F

So what is the television uh I need to explain the few latency for you guys. Kevin is like you guys heard of the P99 gt95.

F

So when we are talking about yeah.

F

That's a HTTP request. It takes, like average, is 100 milliseconds, and sometimes you expect to two seconds. So the P99 is actually the max you take it for the HTTP we said P99 P95 I mean the top five percent request is how long.

F

Latency, which is unexpected from us for Crosby yeah this one, sorry so the these this one. Why.

B

Sorry, the y-axis.

F

Is the normal second of the time taken and then the x-axis is actually the message. The.

D

F

D

F

See there are some spice of it, which is unexpected. So when we say um is the latency of this Square. So this graph is enough to repeat the idea of the students and see okay. So for concurrent concurrent data structure, there will be a problem when we dealing with this okay, when we are doing the factual changes on the concurrent data structure, we will have to either block on the right side or the result. That is confirmed because two trades are actually operating for the same data region to either block the array.

F

To tell the read uh sorry your brother right and then you let the rate operation clear first and then you can start working with or you properly and then keep writing things.

F

So in the interval of writing things you can read so that is, will be some more or something Atomic variables happen in this case the Cross Beam doesn't introduce a lot and it introduces something different which is the copy of the data. So it means writing things and then you can keep reading it and when you are doing the structural changes, let's say there is a tree B3 map you go ahead, you go another level, you are changing the structure of the tree.

F

So what happened?

F

But you keep reading the things and then your pointer changes you and locate things you you need to shop. Your area know this. So what happened? It will copy another copies of the tree and then keep writing it. But don't worry it's keeping within this type of this version.

F

So until this finish and then it will read this one, the second one, the second version, so it would have long. You won't have the latency changes. Then you have to pin the thing. So you, when you are copying things you basically have to manage the garbage collector.

F

um You need to manage the garbage collections of the Hobbies, because.

D

You keep writing things and then English changes a lot.

F

Delete all these copies, so the Cross Beam does have the pin number one to eight and then, when you are dealing with the kingding, you increment the number. So that is how it works. So the.

D

F

Actually caused by the garbage collection, even though you are writing rocks, but you introduce a spy due to the epr a porch base.

F

Reclamation, you reclaim the copies of the data based on the departure based on the version number EBR k. So this is the P4 of this through the structure and then the problem that we face when we are dealing with docules of the state Channel.

F

A PSC Channel stands for multiple producer, single consumer, multiple producer producer message and then you have one reader, keep consuming. It keep dropping the message when you practically okay, we need something to pull the result and the system call. So these two problems actually the pulling how to be proven, and then how do we deal with the system core? So everything happens? So every ring is a very new ways of dealing with the asynchronous, as I stated in the beginning of the talk.

F

So what does the iron urine do? Are you doing? It's actually have two ring popper, which is SQ CQ, which is a short form of a submission q and the condition Q, which is people, do use this short hand to represent these two ring buffer.

F

So for this brings upper when you are dealing with the task, even if you submit a task um on the contrary of the popular is a pool base because I keep talking about result. This one is the submission, you submit a task, and then you pull the pull the data which.

F

And how? How does it work is you're gonna, eat params. This parameters is the parameters to um to create these two things, and then you call this function from the leave. Your urine. You call this thing and then it will occur system call three times is this 3 Ms is the sqcq and the sqe. The sqe is a array of the data which is store all your operations, so NF is a memory map assist, call to tell the colonel to give us memory region.

F

So you use these three memories to create SQ, CQ and sqe, which is the array of the.

F

And then the kernel will maintain the pointer to this CQ so that whenever the kernel finish attacks, it will write things to this CQ and then you are the one responsible to pull the result from it.

F

So how do we submit a tasks? We create an sqe which is the submission Q entry to call this function to submit it, and then what is the operations that support in this one? Basically, there's all this system core yeah, it's like a proceeding, so we wrap the system called with all the parameters, all the arguments it needs, and then you submit to the SQ that the kernel and the details now you already submit to this SQ.

F

You tell the kernel to read it, it's like to put something in it and then you tell the kernel. Okay, you can start to process it. This is the area it will render the kernel.

F

So when you are dealing with the web socket receive when you want to write something right, that's QE will be right. When you try to receive things. Let's say you want to listen to a best soccer connections except get about distributor. Then you submit another sqe, listen received, they keep receiving the data, so this is how we do in ioul rate, a few moments later once the task is finished out. So this is a CQ from the CQ you can prove from it. You can get the condition and then the simple look iterate.

F

This ring buffer taking the cqe, which is the completion Shield entry, taking all these tasks from it, so yeah easy pulling. So you will have a trick to busy pulling pull all this data you're not necessary to cool the color. You don't need to make a system call to the kernel. Tell them hey! When is this task finish? No need just stick from this. One continues.

F

Yeah and then the structure of the cqe in Russ that is quite communicated because I read from the code it's quite complicated and it is a wrap of the C libraries. So right now you have to read from the menu the Linux main page and all this to get the actual cqe here. I am I, compile it for you. So if you look at the main page or the source code of the Linux kernel, it's actually is is that symbol is this and then for the user?

F

Data is actually the unique idea set by the sqe, which is the one who said by you, the user defined one. So when you submit the task, we have to specify the unsign s number to the do it, and then the tasks finish pass back to you, the user, that you basically need to index the tasks that you created and then what? If it is some buffers and all this thing it's not defined by the bro.

E

F

And then for the result of the task, when we repeat it doesn't Define here but will be, then you need to read from the main page.

F

Let's say you need to read from main page because it only gives you the rest a result and a fret so even receive something and the main page for us.

F

He will try to receive something, and then this is the.

F

You base on the flap, you do the buffer select the Fret to get the thing you got in the buffer. What buffer is it? So this is the buffer pointer, so we have to give it a buffer to it. So you need to locate it yourself and pass the pointer to this operation and do basically. This is a system called which is very low level, and then it's not mean to be used by Obama developer.

F

And then for us, there's one wrapper functions in grass that you can do so. It takes the flag and then returns you a number which is the number of the um buffer site. You need to get.

F

So is how we query the result from the kernel, so why is it so important? And then why is it where am I keep emphasizing? It is because the the layout of this IOU ring is actually represented by this graph. So this is the users, as the left side is the user space and then the right size kernel.

F

So whenever you are the one who created you set up, this IOU ring, you actually create SQ and the secret the SQ and the CQ on the user side. You can access it from a user side. Basically, it's a pointer. You can access to it without any locks without any system called thing, but the kernel does have the pointer to this secure and then the allegory instance does help you to manage this memory region, and that is directly writing to this one. So no need to pull.

F

You need to make a system called to prove that result from it and then, when you are looking at this yeah, you can publish the Sysco to the kernel and then the kernel will help you to arrange all the system call it's. Basically a serialization substance is called.

F

You manage all these tasks together in the sequence and then it will help you to finish it and the result, maybe not uh in um family manner, but you can actually pull it from the C2 so far, faster faster sorry. So what do we do? Because you see because we theorize all these and then let the kernel to actually process this or we reduce the contact? Switch?

F

Oh discount CPU to do tasks, you have it's all single traded, so we reduces context switch. We reduce the system. Call frequency got lesser system called the Boom, the result from it, and then we got zero copy of data because the kernel is responsible to write things directly to the CQ.

F

So whenever, when you are dealing with all these websocket connections, you are dealing with the restful and all this you can create multiple iot range.

F

Then you've got two ring buffers you need to iterate. So when there are a lot of IO ring, you basically become the the simulator, which is the legend of king ring. Buffers yeah is that efficient? That is like a job in my company yeah, because I deal with like a lot of exponential growth of the Ring buffer, so I will call myself memory.

E

F

Me, yes, uh you can actually use a command line to to ask to make it as high as you can. Basically, it's just the size.

F

No not only memory size, but also the the file descriptor we made, because you create a lot of descriptor during the initial State yeah, which is you, wouldn't you wouldn't face this problem number. So this is how I got the information. There is one menu written by The, Meta developer, which is formerly Facebook G, is the one that the main author of this IO urine. So he just manage this website, which is quite easy to do with Google. If you type a lot of the io urine, you will get this page.

F

Basically, my idea of the swipe is taken from this, so this is the side very useful. It's like NDM for me whenever you've got a problem with the ilu ringtone processities.

F

Now my stack is not finished. So when I tell my bosses, yes, we already reduce a lot of system call, but unfortunately they are still quite a lot of system. Core happens, it's nearly zero, but it's not zero. Why is it and I tell him that? Okay, because here we receive the thank you, we receive things receive things and then it's only used one. The system code is only used one. So if you want to reuse it, you do submit another call receive.

F

Let's say we need to keep uh keep looping, uh keep receiving our message from a websocket connections. You need to keep submit this receive in the list. I will read, okay, keep on going, keep on going, keep submit and then, whenever you submit it, you need to call this system call. I will ring enter.

F

So my boss tells us why can't you be optimizing up and then I said? uh Actually, yes, so this how I optimize the IOU ring.

G

F

You ring mechanism because I will ring is so flexible. It can help you to optimize things. So these are the three methods that I use: zero copy, Cent, zero Cisco in the whole part and the buffer received.

F

So the first thing is zero copy. So there is one API in the Linux course get together and then in last term is in programming. Language term is the vector IO. So basically you have maintained the I o Vector of these things and then what is it? It is like a pointer to a string.

F

So when you try to send things out, that is a string but you're not necessarily need to pass the string to this um system call. You can pass the pointer to this system called what your string needs to become past multiple string in order to this system call and how does get together system core work. So yeah there is a send message: sexy sexy means zero copy, so we receive the IO priority things, so you can actually pass this. I o Vector to it.

F

So I Vector is the vector of this pointers, so you can actually pass the point of the string. Is there pass it so because you pass it, you need to copy it to the kernel. This first thing. Second thing zeros is called hot part. So yes, when, because the thing.

B

I said you need to.

F

Keep enter the IOU ring to tell the kernel I already submit the task in SQ you're not necessary to do that, because there is a threat called. Are you hearing setup as to pull when you set this wow your CPU, your system.

F

Because that is a background track, helps you to pull the SQ in kernel, so no need to keep calling the SQ uh IO urine enter to tell it will have a spinning Loop to spin in a kernel space keep looping the SQ for you, the current distance, so there's no need to enter okay, and then this one when I use this it's still got the uh I will enter, even though it is almost empty in the SQ. So I look in the menu. Let's see why? Because there is one parameter, called SQ trade idle.

F

This is a timer, so whenever there is no IO urine enter, it will help you to do this, so you will set it large enough. So the are you ring instance will not help you to to call the actor again. So that's what we do to make zeros is called. Another thing is the provided button so parameter buffer is a very new thing, which is happens to be the last year, feature from also from the X I'm, not sure how how to name it, how to spell its name but Expo, okay, explodes Expo!

F

This is the main author of the iot ranking. This um provide us the provided buffer feature so for the password is like you receive things, but now you can get multi-shot receive you receive, so you only needs to submit one receive multiple, so we can keep receiving all this. So no need to tell the kernel to key.

F

uh It's like you, don't need to resubmit the receipt internal. It will help you to do this automatically. So don't you can save more CPU CPU instruction, Co-op right, so it's quite complicated and then I just trained out. But what we got is the QPS increased about six percent and then the latency also drops also I didn't have the Benchmark of the latency. So that's the thing so the user provided buffer is also provided by you.

F

You need to admit yourself and then provide a pointer to this system called call receive multiple to get all these things, so it's also visible in this space Also. So what is the coins of it provided buffer?

F

The concept is, you may be getting batch message because you keep receiving it basically in one notification so notified by the kernel. Oh you pull from the CQ yeah and then you got what you got a batch message so and then it's very low level. So the.

G

First thing: I.

F

F

Because we are using websocket secure, you have the TLs decode it, so you decode it you've got the message: yeah see what I've got when the market is busy. We will get the map the batch message we can see from this, but then you can see some of the uh invented.

F

Why is it that is because of the websocket free the web socket does have the header, so you need to how how much how much length if I forget, is to to slice this into multiple message, and then you need to calculate how much length the verification is to remove the invalid characters. That's how Troublesome it is so yeah.

F

The final part of this coil is actually how are we integrate in the rust, because this is still a very native ways of handling the iOS ring, although rust does have the a single way, but for Tokyo as I say it's a pool base in this system. It's actually submission phase you're, not necessarily need to put things instead, the way you submit things matters so Russ just have some debate in the community, which is the same rate right and the cool stuff.

F

So I think you guys can actually uh read on the comments, because for me, I'm.

E

Not sure about yeah.

F

What is the point of that? The base analysis? Because they are the best talks about how do they prove and bring things? And then.

F

How should I explain to you spending to these types by your IFC Christopher.

C

Comments like they want to get feedbacks on the design. Yes,.

F

They need to get the feedback from all these users and all these the communities, and then they are called their discussion. We actually dictate how we write code in the future, so it is.

F

F

So there are some traits in future. There will be integrating rust to properly uh to should this submission based asynchronous model.

F

So let's see how they, how they do that I can say this, but in the meantime you can only use IOU ring like this. It's me you create things instead of the tricks, because it's not so mature in Tokyo and then for Tokyo in the back end is not necessary. How you read.

F

um I think nostril in other languages.

D

B

Will be more mature.

F

But in terms of the framework and all these things, basically, it's not integrated in the node.js, so you can't see in node.js. Maybe the norm can maybe.

C

Do not have this I can call you later.

F

Because the pattern of how you you submit the same task is different and how docule is.

C

D

C

Like the language support, please.

C

G

F

Will be more mature because.net, they are using the windows asynchronous IO, so there's one I, ocp iocp is also submission based so probably much to that. No need to have big changes in the framework. You know this and then can adapt to this.

G

G

G

G

So it will use, but.

F

um Basically, the doctor is an ABC.

D

Yeah execute director so.

F

Basically, talk show when you want to change it to multiple platform. Let's say you want to make it in a BSC or you want to make it in Windows or Linux. This is the one you need to change, because this one will wait. This executed this for this uh higher level logic, it's still rust, it's not platform dependent, but the reactor is better dependent. So if you change code, it's here but as I said, the the way we cool, reportings and read things added to prove the current result. How is it happen?

F

The logic is not suitable for ioul ring, because I will ring uh it's not proof of data you need to another. Maybe the executed here needs to have maintain the CPU, keep looping and then return to the tasks need to change. Yeah.

C

Yeah right there are two types: always yes and then the other one is completion based I, the four ways we have. We call KQ and IO EPO is the one intimidation, unique space system and windows have their own have their own stuff as well. Here because Tokyo is a the model of the rust facing away, is based on more like polling kind of version.

C

So that's why here we Tokyo uses the if the polling basis systems is the different mechanism which is based on the sufficient base, and the difference is that both ways you keep calling submission, based, disappeared and then once the task is done, you don't you don't tell me you get from another place which is like that system called chain basis is faster and then the point base is, you need to keep sending system call to get the result.

F

And then, even for a single way, a single which syntax uh the using a weight, syntax.

F

Is actually based on the future tasks, so today is one Library called future realize.

F

Yes, so it's based.

D

F

One also so this one does have the book Ready or Not Ready, which you won't see that in iOS, because I will read you got this result. Really it wouldn't be. Not really it won't be done. You will record this result so of the a thing away stuff in Tokyo is built on top of these few Futures. So if you want to reuse it with the hardware, I say that you will not. So if you want to use that you know you can use that feature.

F

And here is my company's name: Arena.

G

B

So now you have to use.

B

And okay, you submit something and you get it then it's done and then you have to request another system call to get the data right. What is the data.

F

Is you need to provide the data? Let's say let's say when we are dealing with the provided buffer? You need to provide this software.

F

How big it is how many slot years so, whenever you receive a thing it will write to this um buffer in one swap one offset it is basically trick is a remarkable. So it gives you this thing and then in the CQ.

F

To get the index of where the data is in your.

F

In your user pointer buffer, so.

D

You inject for me.

F

Pick it up place your data that is complete.

F

Yes, this is how it works, or it is small enough. Let's say it's just a single receive. You can get the data from the pointer so.

F

Is also in the secure it is foreign.

F

They also have a so-called runtime, but the runtime is actually different than the dog show. It also use IOU ring UTS IOU ring, so you can actually write.

E

F

Single way, but in the back end is using I will do a task at you just pull from you. So when you pull from you, yes, you will have a memory limit. So when you use this one, you have to increase the memory limit maximum exercise because it helps you to put inside that. You use a lot of memory to help you to improve things.

F

um To answer. Efficiency basically have another layer to help the proteins in this in this graph. If you have an interpreter layer to do this,.

G

F

So whenever you get the full release, you get the things up.

C

My blue question.

E

E

They are now passed into.

E

Good questions.

F

F

He's a very good interview, questions.

G

At the application, Level.

G

Application there is something that is.

C

C

F

So I think it's also part of the optimizations, where I can solve the tailored, so foreign.

F

I'm still using a.

D

Single connection single.

F

Trade and nothing one trade per connection or definitely will crash one, because you want trade, you use like 4mb of memories. You definitely will run out of memories. You need to solve this up to solve this. Then you see, oh okay, your connections will take a lot of but memories or you take a lot of physical.

F

We keep pulling data yeah, you keep bringing the data to frequent means. So so you cannot handle a lot of users, so users will need to wait for your server to access this, and then you solve it. You optimize by key point which the problem is one by one to solve and softer because chunk of problem first, so.

D

F

Of the where, maybe because, if you what the problem I think there will be that you can change your language say it, you can put credit cash to something else, so you create a database, a lot of optimizations. You need to care.

E

A

I think this is responsible in real time.

A

Computations, for example, three they want to track how many people bought real time like if I go to the website, see now we have 100 000 or they don't sold real time, and the number is legit and those kind of scenario use cases.

C

C

How much faster.

F

um I am not sure about how much faster, because if you want to measure how much faster you have to actually look at the time um the kernel, the kind of fact they will have a time when they receive because I it may be because I, um how do they say this one? So it depends on how I, so the time I observe is when, when I take the result from CQ. Actually it is Maybe seconds.

F

Back then, the kernel right to the CQ so I don't have any way to actually observe that I understand yeah, but we do have um reduce the system call to zero during the program or part. And then we compare neatly compare from the 800 microsecond down to eight micro. Second,.

D

F

D

F

Ten thousand percent.

F

What do they call.

D

B

That lower is the c.

F

Yeah so I think there's a win. I think the big biggest win is actually how we write the code, how we structure the code, that's the biggest win, because this is the PIN form here.

F

Other thing is just a bonus point: it's the biggest point where I'm dealing with a lot of questions and keep creating the vocals trades and background faces.

C

F

I, do I do that or ambitious to to make it as one but firing you've got a lot of system called there right. You need to index all the have been available.

F

F

B

F

Then we identify the user data from the CQ. We.

B

F

Because when I want to do the restful, I also use this I will ring so I will mess up the better than I just create two Lu ring distance one to handle all these website connections, one to handle all the restful connection. That's cool and notice. It's better for me because.

F

F

Creating more is cheaper and more easy to make.

G

Any other questions.

F

At this one right, so this one targeted easy. It seems like we created all this red.

F

This task, but when you do the OS trade, if you also always does have a scheduler just depends on you want the document for you or you want the osupender for you, but always is more expensive, because the West will schedule CPUs. You can do all these contacts. Okay,.

E

We do that I mean.

F

I, don't see that so.

E

F

I mean the main point you want to reduce. This is yes, the main point is reduce assist call, but in here we are not reducing the Cisco from Tokyo shoes to our urine is actually we reduce the contact switch because for this one of your spawn the doctor to this task, it will somehow pull from.

B

F

State and then you come to this task, he will also pull from this. So when you start switching this task, it will have a contact switch. Also now what I say is civil rights. The system coins, because we using one Loop to handle all physical to reduce the condensation.

C

So Tokyo runs on refresh the Russia we practice and then for the Tokyo. This is Tokyo spawn docking. There.

C

A

Any other question the question then hey. Okay, so uh thank you. Foreign.