WEBVTT 00:00.000 --> 00:10.000 Now, you should see my slides. Cool. Yeah, thank you. Thank you for your kind introduction. 00:10.000 --> 00:17.760 Yeah. You were always told, I'm showing you about skatex, and I've stick a slider, so 00:17.760 --> 00:23.200 you can teach us whatever. Anyway, let's get started. You probably all know that we have 00:23.200 --> 00:28.240 multiple CPUs, and I always assume for example, in this system, I have 8 CPUs, but 00:28.240 --> 00:34.560 I have multiple tasks. For example, I've fully like 8 IDs or something running, and so 00:34.560 --> 00:40.240 what we now want is we want to scale release on the CPU. To make it easier, we say we have 00:40.240 --> 00:44.960 only like 2 tasks, and we have only 1 CPUs, like in the older days. And what we want to do, 00:44.960 --> 00:50.960 we want to run both applications of both tasks. These tasks might be purchased. They might be 00:50.960 --> 00:59.120 referred to or what else. And so what we could do, we could just run task 1 all the time, 00:59.120 --> 01:03.440 but this won't be fair. We will be starving task 2 and everyone will be great. So what we 01:03.440 --> 01:08.800 usually do, we scale task 1 for a bit, we scale task 2 for a bit, then we are like a task 1, 01:08.800 --> 01:12.720 and we scale with it, but then task 2 doesn't want to come, because it's probably 01:12.720 --> 01:18.320 writing for a database connection, so we scale task 1 again, and so on. And that's essentially 01:18.320 --> 01:23.680 called scheduling. The thing is scaling is really nice, and everything, but the only problem is, 01:23.680 --> 01:29.040 it's in the older days, like before Linux 6.12, it was really hard to create your own, 01:29.040 --> 01:35.120 because as someone told me, within the now we taught my life here, it was like 3, 4 people in the 01:35.120 --> 01:41.680 whole world knew how to do it, it was a cane code. That was pretty much a problem, because we 01:41.680 --> 01:49.120 started innovation with this. And so what's nowadays is that we can create our own using 01:49.120 --> 01:54.960 the scalar extensions and EVPF. You probably all know about EVPF here, so I won't be 01:54.960 --> 02:00.480 going into details, but I tell you now about schedule a bit. So what is schedule, it's essentially 02:00.480 --> 02:06.640 the scalar extensions of the Linux code, which means these extensions made to EVPF like you, 02:06.640 --> 02:13.920 for example, know the XCP subsystem, which you can write, package filters, or other subsystems 02:13.920 --> 02:18.800 like the Linux security model, this is essentially the same bad for scheduling. The person that you 02:18.800 --> 02:25.120 see here with his halo, his vision, he wrote, tell us later, getting this in, and I quote him 02:25.120 --> 02:32.080 here with, or he thought was the most important stuff about analytics. He said, the benefits are 02:32.080 --> 02:38.000 of one, the ease of experimentation. I can show you here on the slides of full working 02:38.000 --> 02:43.360 scheduler that can run a whole presentation. It allows us to customize, because we can write 02:43.360 --> 02:48.800 our own scheduler, we're not stuck with like EVDF, that's in this code, it's called, but 02:48.800 --> 02:53.600 there are some instance where we might want to have a different scheduler, and it also allows us 02:53.600 --> 02:58.800 the rapid deployment of schedulers. The main thing is, as someone told me, who did some 02:58.800 --> 03:04.240 call involvement with like in the olden days when we wanted to write a scheduler, and deployed, 03:04.240 --> 03:10.480 we had to restart like 1000 machines, and then wait, of course, before we had to recompile a comma, 03:10.480 --> 03:14.800 everything, and that's horrible. So we get all the benefits of EVPF in the scheduler world, 03:14.800 --> 03:19.600 and that's pretty cool, and it allows us to make rapid progress, as many of you have 03:19.600 --> 03:25.600 probably seen because like the BVF scaler are beating scaler in the gaming world, for example, 03:25.600 --> 03:31.840 and it's pretty nice. So I want to show you the simple schedule dance. It's essentially how 03:31.840 --> 03:37.760 a basic scheduler works. More than more complex scalers have more things, but this is like the 03:37.760 --> 03:43.040 most basic version. So on top of our task day come in, and then we have essentially two kinds 03:43.040 --> 03:49.600 of cues. We have global cues that our scheduler uses to have state. These can be priority cues 03:49.760 --> 03:55.120 are usually day addressed, simple cues, as you know them from computer science classes. 03:55.120 --> 04:00.480 And then we have local cue scenes, local cues are fine to CPUs, so every CPU has a 04:01.680 --> 04:06.960 local key on what we do when the task comes free. It's like going into the trouble too, 04:06.960 --> 04:12.880 usually, and then every time a CPU is like, I need some work, or we tell the CPU, 04:12.880 --> 04:19.360 you need some work, drop everything you do, it's called kicking. So we literally kick our CPUs, 04:19.360 --> 04:24.160 which is kind of implored, but so what we do when the CPU asks us, I'll say, 04:25.200 --> 04:30.720 give me a task, and we're like, we're using our totally cool new algorithms, 04:30.720 --> 04:36.320 ask the global cue from the element, and then we'll find. And then when the CPU is like finished, 04:36.320 --> 04:40.720 and it's like with the task, because task run in a specific time size, as I showed you in the 04:40.720 --> 04:46.640 beginning, then the task goes back into a global cue. And essentially these distance, 04:46.640 --> 04:52.000 how we cope with the choreography of this dance, this is what we do with scale extensions. 04:52.000 --> 04:56.240 And the cool thing is, it's highly abstracted, so you don't have to implement the project. You 04:56.240 --> 05:01.680 don't have to implement all the medically details of how the signals really can scaling, 05:01.680 --> 05:08.080 which makes it usable. And even the best thing here is that whenever one task is stuck in 05:08.080 --> 05:13.920 not executing these parts, it's gone for like 30 seconds, then Linux is like 30 seconds is almost 05:13.920 --> 05:19.200 infinity and scaling. I just kick your scheduler, and it's really nice, and I can't tell you how 05:19.200 --> 05:23.360 often this happened with this machine, and it's cool, because don't have to restart the system. 05:23.360 --> 05:27.840 After 30 seconds, even if you do some creation mistake, it's back mobile, and it's cool for 05:27.840 --> 05:34.080 life timers, I can't tell you. As we see in a minute. Anyway, so essentially how our basic 05:34.080 --> 05:39.280 serial rigs is, we have here our NT work, that's essentially called whenever task is like, 05:39.280 --> 05:45.520 I'm ready to be pushed on, then we have our main scale link to you, and then essentially we move 05:45.520 --> 05:52.080 it to the local queue, and then we have an NQR, the NQR is also called when the task is like, 05:52.080 --> 05:57.200 finished with its time slides. Be aware that there have been name changes, so the names 05:57.200 --> 06:01.840 they show here are the 6, the 30 names. They were changed, because the openings were really 06:01.840 --> 06:06.240 confusing, because there were two things that were called dispatch. I can tell you, that's not 06:06.240 --> 06:12.800 great, and then we change it. That's cool, and my code in this, it's a really cool community, 06:12.800 --> 06:17.680 it's something that's confusing, just follow them, and in the end, you see your name on the 06:17.680 --> 06:23.600 local mail list. Anyway, so these are the main changes, I don't go over them, it's just when 06:23.600 --> 06:28.800 you see existing examples, but the openings work, because as we know, EPPF is like stable and such, 06:28.800 --> 06:35.280 so how to say now to our scale loss, a basic cell law works, a basic title scale law works 06:35.280 --> 06:40.240 essentially by we get tasks in, and the same all that they get, and we take them out. It's pretty 06:40.240 --> 06:46.800 simple, it has advantages, because we don't, we scale in, in like a deterministic fashion, 06:47.520 --> 06:55.360 but there are also another print, as we probably see now, if anything goes correct. So here you see, 06:55.440 --> 07:02.080 he's a simple schedule, before he's just some, some definition, some definitions that we need, 07:02.080 --> 07:10.080 but they are not important, so what we, what we start with, we start with creating our scale 07:10.080 --> 07:17.200 on Q, so in the end it method, we create our scale on Q and we can then use it as our global 07:17.200 --> 07:21.760 Q, and what we don't do, we have to end Q, so totally that's when the task is like, 07:21.760 --> 07:26.400 ready, and it's like, I want to be scheduled, please, and then what we do, we compute the 07:26.400 --> 07:32.560 slices like the amount of time that the process can run at the time, and then we insert into a 07:32.560 --> 07:38.480 local Q, here we just, and then we insert it into the global Q, so totally before, and then take 07:38.480 --> 07:43.920 it out later. The cool thing is here the slices in, in micro seconds, and I, in nano seconds, 07:43.920 --> 07:48.720 and I can show you later what happens when we make it a little bit larger, and you can see the 07:48.720 --> 07:54.400 direct effect, what we usually want to do is, and what makes the scale of, somewhat usable 07:54.400 --> 07:58.880 in the direct to seconds, that the longer the Q is the, the smaller our time slices become, 07:59.600 --> 08:04.320 it just makes it so that when our system gets overloaded, we can still schedule stuff, 08:04.320 --> 08:09.440 and then, um, to this point, it's pretty simple, which is so like, pet the fruit's task, 08:09.440 --> 08:13.280 I don't care, and it's just more than this, so for example, it takes care, 08:14.160 --> 08:19.760 you know, that maybe some process can't be run on all, some tasks can be run on all CPUs, 08:20.880 --> 08:25.600 and it takes cage of this, and it's pretty cool, because cave threads are slightly different, 08:25.600 --> 08:32.800 as you see in a minute, so that works. Let's see, it's a live demo, it's a live demo, wish me luck. 08:32.880 --> 08:41.760 Um, thank you, um, the major task for presentation, remember in one's password, 08:46.240 --> 08:52.960 I think that's like, that that's like the major task, so, no, I'm imploshing, it will get worse, 08:53.840 --> 09:05.360 it's quite embarrassing, but a bit, so, um, essentially, um, it's, it's essentially, 09:05.360 --> 09:11.760 I have to schedule our attention, it's cool, and what we can, to look at, what skin law we have, 09:11.760 --> 09:19.360 we can just say, um, you see, to cut, get a fruit up, so we can, we can look up the name, 09:19.360 --> 09:24.320 and we define the name here later, so it's just struck up, so we define the name here, 09:24.320 --> 09:28.800 and that's pretty simple, so we can have a schedule, as you see, maybe the demos is slightly, 09:30.400 --> 09:36.000 it could be a little bit more responsive, so if we add to your service, so our time source is 09:36.000 --> 09:47.120 now 50 milliseconds, don't do this at home, kids, um, start, now it's the out, yeah, I have to 09:48.080 --> 09:53.760 no, it should work, yeah, it's still the minimum skill, like it, now it's getting a little bit more 09:53.760 --> 10:03.200 choppy, um, and you can use this, and G, with like two cores, there should be some sloppiness, 10:03.200 --> 10:10.880 anyway, um, they is trust me, um, you see here, that's a pretty basic skill, and that's really cool, 10:10.880 --> 10:16.400 that you have like 10 lines of C-code, and you have a stellar, now on to more interesting 10:16.400 --> 10:20.960 servers, um, this is one scale that I really like when I was in operating system class, 10:21.760 --> 10:28.160 because it's called a lottery scale, so then she, um, like a lottery, you pull, um, you pull, 10:28.720 --> 10:34.320 you put all the tasks that you have in, in your queue, into a ball, in, in, like a lottery ball, 10:34.400 --> 10:41.040 and then from randomly from it, and that's nice, because, um, it gives us non, non-determinism, 10:41.040 --> 10:46.320 we can use this ball, for like a maybe in ranges of cryptography, I don't know, and for testing, 10:46.320 --> 10:49.520 if you want to know about how we can use non-determinism, and testing, 10:49.520 --> 10:54.160 there was a talk in the testing that room this morning, um, I can really recommend it, 10:54.720 --> 11:00.160 because it was mine, or by the person's alley there, because that only came after the talk was finished, 11:00.480 --> 11:06.240 um, trains are also under-deterministic. Anyway, um, how do you implement a lottery scale? 11:06.240 --> 11:09.920 It's pretty simple, let's get in it, as you see here, it's the same, 11:10.560 --> 11:14.880 the scattering queue is also the same, because we're also putting it in a cloud book here, 11:14.880 --> 11:20.240 the only difference is here. Um, it might seem like a lot of code, but it isn't, so essentially, 11:20.240 --> 11:25.600 what we do here, we, we loop, we can loop over the queue. We can, in, in, in this line, we can loop 11:25.600 --> 11:29.040 over the queue, and essentially what we're doing, we're first taking us to run the number, 11:29.040 --> 11:33.600 and then we like take a random number between like the, the start of between a zero, 11:33.600 --> 11:39.760 and the end of the queue, and then it's just walked away, got there, and also you see a BTF CPU 11:39.760 --> 11:46.000 mask has, we, we just sit more if we can't run it there and like skip it, because sometimes 11:46.000 --> 11:50.640 it's that like all flights, they can only be run on certain CPUs, and then we just switch it, 11:50.640 --> 11:55.840 and the important here in this example is I added a print, so I can show you that it really works, 11:55.920 --> 12:06.880 so let's, let's look at it, start as H lottery, now, and if everything goes well, 12:07.840 --> 12:13.280 you see here, um, well, it's still all of us. It tells us exactly, because it's just that 12:13.280 --> 12:18.320 BTF in case, you probably all know, like, for logging, and you see here, where we're 12:18.320 --> 12:23.680 selling the multicolored scaling cable, because the system is quite idle, when we move the 12:23.760 --> 12:29.120 cruiser, we see something more, and that's pretty cool, because we've now written a 12:29.120 --> 12:34.480 scale that locks us every scale length decision, we can insert BTF in case everywhere, we could 12:34.480 --> 12:42.080 even like, um, recorded to output sound and anything, I've done weird stuff, um, if you have 12:42.080 --> 12:50.880 other weird ideas, wait for it to all the end, um, so that's that's all, um, so of course, um, 12:50.960 --> 12:57.600 faster than I expected, but anyway, so of course, um, you don't have to write C code, 12:58.640 --> 13:04.480 um, you can also write Java code, because with a simple, with a simple, because with a simple, 13:05.520 --> 13:11.520 implement schedule, you have a scheduler, and in Rust, it's also, uh, comparably easy, so 13:11.520 --> 13:17.120 you don't have to write C code all the time, and what you also don't have to do, you don't have to, 13:17.200 --> 13:21.920 like, manually, use BTF tool, like, I did, um, if you want to know more of these, 13:21.920 --> 13:25.520 like, to correct code for this, and the cool thing with, well, for if the scale, like, 13:25.520 --> 13:30.320 sentient that it opens to door for, for other people to work with skills, because before it 13:30.320 --> 13:37.040 works like, the, the reclusive thing of, like, a few nodes, and now we can show this to people, 13:37.040 --> 13:41.600 and be like, and I can show you, like, write your own scheduler, like, it's not hard, 13:41.600 --> 13:46.240 and it's a cool thing with EVPF, because the choice of words at all, it is. But so, 13:46.320 --> 13:52.480 pretty small group of people that really distinguish it at a conference, ask us about scale X, 13:52.480 --> 13:58.560 when you see them, um, and that's pretty cool, and, and nothing they want to mention, 13:58.560 --> 14:04.320 during the, during the work and this, um, I felt that verify issues can be quite nasty, 14:04.320 --> 14:10.160 so it started a verify error project, so if you want to contribute, contribute your verify errors, 14:10.160 --> 14:15.360 because in the end I think, um, what we want to do with EVPF is, make it also 14:15.360 --> 14:20.720 educational to it, because before we start learning and scale in, in, like, operating system 14:20.720 --> 14:25.520 classes, we told them like, oh, scale, scale, scale, it's really hard, it's like, 14:25.520 --> 14:29.600 no, you can't implement your own, but now we can be like, here are 10 lines of sea code, 14:29.600 --> 14:33.760 and you have your scheduler, and that's, and that's very cool and do experiments on it. 14:33.760 --> 14:39.440 And, now I'm coming to the end, um, the thing that, that we launched just for customers, 14:39.440 --> 14:44.880 you see the t-shirt, and you see this mark. The thing is, we, we made a scheduler context, 14:44.880 --> 14:50.080 we, that's like, all the people behind scatex, you can win a mark, and the t-shirt, 14:50.080 --> 14:55.200 if you submit your own scheduler, um, so you can either submit your best-scaling ideas that 14:55.200 --> 15:00.640 gets you t-shirt, or you can submit the tiny implementation of a scheduler, whatever you want 15:00.640 --> 15:06.640 to do with the scatex, we're open to any language, program it in batch, if you write the 15:06.720 --> 15:12.800 library for this. But just submit it here, and, um, you have time till tomorrow at 10, 15:12.800 --> 15:18.560 and then be able to give the prizes to the best, whatever best means, that's, there are 15:18.560 --> 15:25.200 our interpretation, um, in, um, trail release talk, in the, uh, Linux, in the, I think it was the 15:25.200 --> 15:32.160 code that room, you can find all the informations here in this repository, and I hope you submit something. 15:32.240 --> 15:37.440 I was Johannes Bechtberger, you find my blog at mostly another study, but write random stuff, 15:37.440 --> 15:44.560 counting most scaling, um, and you can find more about, uh, my team, uh, which is submission here, 15:44.560 --> 15:51.120 and the QR code should lead you to the scheduler context. And I think that's all here. Thank you. 15:51.120 --> 15:53.120 Thank you. 15:57.120 --> 15:59.120 Do we have time for questions? 16:01.120 --> 16:03.120 Yes. 16:22.080 --> 16:28.000 Yes. So the question was, is, is, is, is, is it, it could be a real benefit for real-time 16:28.000 --> 16:32.720 comments, these custom scalars? This custom scalars are a benefit for everything that has 16:32.720 --> 16:37.520 dominated, where we know more than like the schedule developers, which are many applications, 16:37.520 --> 16:42.080 for example, that the companies developing their own scalars to improve gaming, to improve the 16:42.080 --> 16:48.720 server performance of websites that you use daily, because if they can like, inch out, get out, 16:48.720 --> 16:53.440 like, half a percent of performance, that's really a lot if you have a couple of hundreds of 16:53.440 --> 16:56.720 thousands of service. Um, yeah. 16:56.720 --> 16:58.720 Good for a lot of questions. Oh, even one more, yes. 16:58.720 --> 17:02.720 Thank you for the invitation. Thanks for that energy. 17:02.720 --> 17:09.760 I have anybody's right, sure, like, free hands, uh, the existing schedule just to see how much 17:09.760 --> 17:16.080 I covered the screen. So the, the question was, could we could implement existing scalars with this? 17:16.160 --> 17:22.000 The thing is probably, but it doesn't make any sense for honest. Uh, yeah, to check the overheads, 17:22.000 --> 17:27.120 yes, it could make sense, but the thing is in the end, the overhead is, uh, negligible, 17:27.120 --> 17:32.160 because scheduling is never like the thing that costs you performance, that she is going 17:32.160 --> 17:36.240 to say, and we want to make them as fast as possible, and typically we want to implement 17:36.240 --> 17:41.920 new cool scheduling decisions. So most of the workers focus there, you find more and you find 17:41.920 --> 17:47.360 more in this in like the schedule of contest. So please submit within have any submission. 17:47.360 --> 17:52.080 So your chances are really high together. Nice, you should. And this is exactly how.