VideoPoints

• ••• • •• • ••••• • •
COSC 6365 Introduction To High Performance Computing - Lecture01_Course_Intro
Help Tour
© Distribution of this video is restricted by its owner
Transcript ×
Auto highlight
Font-size
00:03 all right. I heard that skin . Hopefully, that's okay with
00:08 So I guess I should start to myself. So, as you
00:14 in fact, in the computer science and what you may not know is
00:19 the guests have been interesting high performance course forever. Pretty much ever since
00:26 graduated, um, and worked a of years and industry as well.
00:34 that will be reflected in some. I approach is class in the top
00:42 and off the years. And industry probably about half a member and as
00:49 type in the strict on the other was in computer. Look, I
00:56 tennis years in each. So in of introduction, that's probably enough for
01:04 . And so yes, Um, . My PhD student. Let him
01:11 himself in a minute. Uh, that the class is too small to
01:16 having an official ta. So Yeah. Kind enough to volunteer.
01:24 , help with the class unit. look, issue an Arctic sea for
01:28 class. Hey, what? I to talk about background and what you
01:36 . Eso Johnson said I'm from a student in my third year of fiercely
01:43 , after Johnson, my research focuses high performance and energy efficient computing.
01:53 , as promised, the scores goes for the price now. So that
02:02 , I have some experience with the and on the programs that would be
02:07 and the platforms that will be using on. Yeah, so just reach
02:13 to me for any questions you might . Okay, so, um,
02:25 and while trying to first just go a little bit but, um,
02:30 to happen in the class and then a little bit about the background.
02:35 , I think you should be interested taking the class. So first,
02:39 just kind of just showing what this the syllabus. It's nothing. The
02:48 what it's already said. They're the so being for emphasis on the
02:58 as is on the edge of the HPC high performance computing. So it's
03:07 emphasizing, trying to get high performance off whatever platform you maybe using later
03:18 in your life. That's the So it's kind of on resource utilization
03:26 techniques for doing that. No, organization off the courses and that the
03:36 so electricity is tried to familiarize to both the environment in which contravene using
03:46 for assignments and projects. And that's likely come to be. They fared
03:53 . The shared environment is what way from the class and junior have access
03:58 other environments. Do that from You're welcome to use us long
04:03 It serves the purpose of assignments of . But we put, um,
04:10 us a set of shared environment that new to some student and often to
04:18 teas, and it is set up an intruder. Courson. For
04:26 I'm trying to give youth in understanding what? Um, happens what's available
04:34 in most environments. Ah, and that May is, uh, multi
04:43 processors is over to stay. The thing that is out there your love
04:47 even in your phone. Um, than there are people use more on
04:58 that's class of accelerators of That's kind what's known as the true genius
05:04 And then they're clusters. And, , well, you may be familiar
05:09 all of them or not. this is not a single whale
05:18 so I don hear health, falling from sitting for some friends.
05:31 , but so, unfortunately, the environment is not the same across all
05:37 platforms and soul. There is particular that will get exposure to in the
05:43 . The open MP on that is kind of homogeneous, no programming in
05:51 . What has evolved even now support genius knows open a sissy bus designed
05:59 Head Virginia's knows. But both of two are basically no level programming.
06:05 then when it comes to clusters, what's known as MP I or Mrs
06:10 into face. It's so that will part of your got to do.
06:21 would say one assignment primarily on each of these three programming paradigms. E
06:32 , there will be a sister one one of these, uh, programming
06:38 . And then there will be initially that I just came to get too
06:44 with environments, and the tools should to understand the performance of their
06:50 Ah, and, um, the is in order to get performance,
06:58 really necessary. Teoh understand the platforms using and their properties. So because
07:07 programs has kind of a matter mapping applications onto the particular hardware platforms
07:14 unfortunately, compilers or not all this capable of resolving the difference with how
07:24 describe your problem in a programming language then what other things needs to be
07:31 to execute? Well, uh, a different platform. So that's why
07:40 also in the beginning of the Let's there will be a couple of
07:45 I talked about the architecture at more a systems level. That is you
07:53 importance for the applications that on So you kind of get some insight
08:01 how you structure cold so they would well on the part forms that are
08:11 . Ah, something convention. Early , there will be environment we
08:17 Hopefully, I haven't gotten approved yet that it has been approved. Every
08:22 were effort to focus on a National Foundation and funded center in Pittsburgh numbness
08:33 per supercomputing center. They are part a go national collaboration owners exceeds.
08:41 exceed is kind of the framework in is used to computers. Uh,
08:47 , and we'll see. We may to use another center to that provided
08:53 in the past. Most of the , he just use this,
08:58 this person pickle giving center for assignments they have both. Typical just a
09:08 . The processor knows, as in clusters as well, the complete
09:13 But they also have GP use on notes. So it's an environment that
09:18 shoes a chance to test both low programming hatred, gayness, programming in
09:24 program. Uh, the grating is , Cares is, could wait in
09:33 total score, which is ah, on the great will just come from
09:44 assignments. And then there's a midterm that has been to take home determines
09:51 a little bit kind of obvious, guess in things case, because everything
09:57 online at this point. But you it will have a week to do
10:04 midterm, and there is a final instead of a fine an exam.
10:12 , saying hope it's on intent is there will be three or four week
10:19 . So that's why it's carrots affair . Three. Wait and,
10:25 the total score. Ah ah Try to keep you at pay so
10:35 don't fall behind. Part of that that look kind of penalize late
10:44 But if their issues with a so it's not something you cannot
10:49 There were obviously not penalize late If there are things beyond your control
11:02 ah, thank you L a prerequisite Central from familiarity with programming and the
11:12 types environments, um, the programs nuts you will not be asked to
11:19 a lot of coding. It will mostly provided colds, and you may
11:24 to at something or correct something, to make it work properly or try
11:31 optimize the cold is given to so the action coding will not be
11:38 extensive. But of course, you to have some understanding her programming languages
11:45 as well as on the UNIX A little Leonard for some flavor
11:54 So Electra material. I don't have book them for a reason why the
11:59 is I am. It's on the . That action covers the material and
12:05 breath. We're doing it so on special things for algorithms or particular things
12:11 you program and you can find But in trying to do kind of
12:17 broader interest embassy, then I haven't any books that it will be electrics
12:28 Supposed to come back Born Andreja Now we're post on the BU point
12:36 old website on and, uh, to post them as soon as at
12:44 ? This you takes a little bit the lecture is over into converted to
12:48 before. So you should get posted same day. The videos and as
12:56 think I mentioned, see ashes sopping on his own kind time on your
13:05 information. So that missy ones next anything No, I was I thought
13:13 was yes. I thought it wasn't less so. So the emphasis Issa
13:25 IHS, to understand how you get performance out of colts and to do
13:32 , but needs to understand what the is that you're actually using. So
13:38 that again, there will be through lectures based clearly trying to tell you
13:45 the important features are. Neither processors keep use or clusters that have a
13:57 the impact on performance. Um, I mentioned, there will be,
14:06 know, yes. One assignments. will not become expert on any one
14:10 the program and paradigms, but you get the sensible thing are and how
14:15 use them. Oh, Ah ah also said we'll have tools. We
14:23 to figure out how she wants a performance while they need to figure no
14:28 actually assess it on the way it is to you some fools.
14:36 so there's like this a couple of on the tools again in the
14:41 So I understand how to says Um, one thing that is common
14:54 it's a threat to you, of , is the The time is the
15:04 critical part is how to deal with data and what's called the data
15:12 Some of this case, it has do what parts of the system is
15:17 on. The memory assisted me slow it comes to clusters, networks of
15:22 so understanding impact off how you access work with the data. That is
15:31 you're Houston for. The competition is to get in good performance, and
15:39 this is the ice orbit. It's . So for basically two for you
15:46 see best that they are a sign classes. Sin friggin set the expectation
15:52 when things will happen and what things happen and the kind of assignment scheduled
15:58 will have for the class. so I think, uh,
16:09 I can show this slide and then . So yes, perhaps at some
16:15 in terms, off experience, both ? It from the time taking the
16:22 himself as well. That's the The price before. But we're trying
16:27 give you an assignment which is supposed get yourself an account on that
16:34 You can ask request even before the or compute cycles is given to the
16:41 . But, um, so you're , Teoh, go to the exceed
16:48 on fill in what? You need get yourself on account. Then once
16:55 is an allocation Oh, resource is the class. Then I will,
17:01 huge to the class. So you to use the class resources, but
17:06 need to have your own account, , in order to being able to
17:12 the resources. So stop there for minute and first asked Yussuf he has
17:20 . And then and I'm one of . Certainly free. You're free to
17:25 . Ask questions and then we can . So for Johnson, mentioned in
17:36 of assignments, will be using the's platforms. So you get your account
17:46 it. If you already have an on exceed, send us your user
17:51 and they love you to the class allocation, uh, in terms off
17:58 assignments again. What? People need . So we will provide the codes
18:06 the most part. Mostly be off old and see eso. But as
18:14 believe you mentioned that people need to make slight modifications to the cold.
18:21 everything done that's being asked in the assignments. Um, the first few
18:28 be to getting familiarized with these units and Sofia configuration of these notes are
18:38 , But all the specifications are off machines that are on the first
18:43 They feel the benchmarks like a generally in HP speaking, get the performance
18:52 performance. So to say, off machines will get you coming. The
18:57 with that, uh, as faras submissions goes. So since I believe
19:07 Johnson can save much better, uh, we expect assignments to be
19:16 you to understand how the system works than just running a bunch off course
19:23 monitoring the performance of what that means we provide you with the poles,
19:28 run them. And based on the to you provide your insights that what
19:34 think is going on by the performance high, right of performances lower
19:40 the memory bandwidth low and so So it's more off how you get
19:48 understand how this clipper, how these for rather than just performing a bunch
19:53 programming exercises, that's it. So anyone feel free to ask questions.
20:06 fair question your sneaker. Remember to yourself. Um, so we can
20:14 you. There are questions. Thank you. Can also use chat
20:29 if I asked them in training for moment with a chance. So,
20:38 OK, so I think that switch , what a little bit.
20:52 substance in the class will give you . So today's simply trying to motivate
21:00 . Um, what you will learn where the what you will learn where
21:09 someone you may already have, very clear. Goes for what?
21:15 have to use what you learn. about this man? Enough Have the
21:20 perspective. So I'll try to give perspective and again probably comes to a
21:26 extent from the fact that I worked industry planned great many years, either
21:30 is busy or designing software from Or designing computer systems for his
21:39 So one thing that I I would it started a Miss Norm er and
21:50 coming from, I think the way cuisine. Something started that there waas
21:58 in science and to some degree engineering there. Waas. So it be
22:04 many it became synonymous with certain aspects scientific engineering computation. I would say
22:15 it's important in a very large number areas, and I'll try today to
22:22 why it's important in abroad this matter it's not an application domain, but
22:30 actually a set of methodologies and techniques is very useful across the very broad
22:38 off applications over businesses or however you to look at it. And that's
22:45 I was going to try to convey you today. Hello, So as
22:50 mentioned, Student originally was really science physics chemistry and later on in life
22:59 . Um, so they're very for of really just for science on Ben
23:07 started. Teoh uses today is widely for defining all kinds of products,
23:13 I'll give you examples of that. , more recently, Data Analytics.
23:21 a commercial or business aspects of not product design has started to use
23:29 , for which the high performance is important aspect. And, I would
23:36 most recently with all the different both networks appearing and sensors in scientific experiments
23:45 tons of data. The real time have also become a critical court or
23:51 domain for which SPC is applicable in of methodologies. So this is a
23:58 bit outlined. What I don't try Dr Box to give you some background
24:07 what's happening or where is busy is . So yes, very quickly go
24:15 a number off applications not to try teach. And it's much a specific
24:23 first example. If I applications where is critical and at the same time
24:30 you some sense off, uh, scale or the challenges in those
24:38 as is the reason why performance is if there is, if you can
24:47 things on your on a leisurely then performance may not be so
24:54 and but the examples I'll give stress so Healthcare first. So this is
25:02 that some of you may have heard . It has had mixed success.
25:08 was I was a hi Piston is of a bad word. That was
25:16 of excitement a few years back. , about what? The on call
25:22 Watson. That was basically computer That's where and designed to how in
25:30 treatments for sick people. And they to deploy it, and they did
25:38 it on as they I guess, many times discovered that the challenges was
25:45 little bit harder than they had So it's not that it was a
25:51 , that it didn't quite live. , I and it was basically in
25:58 tight race systems. And here's a bit, some more data points of
26:03 . And given that I don't have book, so I have slides with
26:08 little bit too much text to be in the presentation. But that's really
26:14 give you some documentation. Oh, in case you're interested both in terms
26:20 getting quantitative data as well as some to where you can find more information
26:29 particular area where application is of interest you? Um, no. Mammography
26:39 another one in terms of health And put this in here because it's
26:44 common use. And it does produce fair amount of data on that will
26:49 back to talking. But data and Analytics and big Data has been,
26:57 , up for a number of You know, right now, during
27:02 pandemic, Data Analytics seems to have a little bit to the back
27:06 In general, it's the question useful information on a very large data
27:13 no in terms over rial time. given were in Houston. And this
27:21 out for the oil industry increasingly uses time monitoring our wells and pipelines and
27:30 things in order to, uh, things more efficient and save money.
27:37 it's just you can see from this there is this live on the money
27:41 in this decree. If I'm this is another one that does both
27:46 time as well as design and post . That kind of an interesting.
27:54 terms of green energy and windmills, or technology is used not only in
28:02 the windmills, but it's also used figuring out where to place the
28:07 so they look at weather patterns and fears of times and locations that are
28:17 and placements for when Bill farms. so they helped potential customers figuring out
28:25 is a good place to put the then once they're in place, they
28:29 do it for the other control of in the windmills. And don't go
28:35 all kinds off teachers in how you the wings of the and orientation win
28:46 other things that is, you monitoring. Ah, turbines. Turbines
28:51 up in jet engines that shows up part plants on and for generating
28:59 And this is about the jet and they are continuously monitored and
29:07 Ah, a lot of data also the design off turbines. So,
29:13 it says in this one, um, yes, for designing a
29:20 and then j produce about petabytes of . So it's again just to give
29:26 an idea. Else came. It's the case that get engines are continuously
29:31 when they are in operation, and there are no left different manufacturers for
29:39 Electric or, uh, Rolls That bill Jeff Daniels. Do They
29:49 centrists of monitors, their engines and to tell the airlines if there are
29:56 that they need maintenance to try to it so you don't have to wait
30:01 something goes bad on that to get of time. So the downtime off
30:09 gets minimized. My hiring basically tried be preventive maintenance plants. Um,
30:18 one is just from Airbus. you're being manufacturer that's similar to
30:26 Just happened to get right from the folks, but it's also isn't
30:32 You can get an idea off amount of data that is being produced
30:40 monitoring and assisted this 1 May the something current person your plans today for
30:47 , Katie. I think there were 10,000 sensors Ah, in the
30:54 And many of them are monitored in time. And then they know kinds
30:58 data analysis analytics on the later something they think of as being, um
31:06 all technology and totally boring, like on. So this is from Union
31:13 and freight trains. But they do lot of monitoring in order to has
31:19 say in this particular slight to try prevent derailing from Well, is going
31:25 or tracks going bad? So I some on here is you know,
31:32 only thing autonomous vehicles That is another of our thing these days that and
31:38 produce a lot of real time and I'll come back to that later
31:45 . So this is a few So we're, um real time monitoring
31:50 important towards you. Have enough Theda, Not you need to work
31:55 Teoh require it's high performance out of platforms you were actually using to keep
32:02 with the data stream. I'm getting example Amore in the engineering the May
32:09 so much real time. But this is for with basically exploration that again
32:17 in Houston and energy type application on in this case is best give to
32:26 . And what's the surveys of the oh areas for oil exploration?
32:34 what the data sets may be but also the level of processing.
32:39 this is a few strategy is one the oil companies. Um, it's
32:44 French one, But some of this the process about 100 pounder by
32:49 a few years back and you have other some just example from BP what
32:56 did in terms of this, I time exploration work. Uh, I
33:01 this. I don't have any particular associated with this one, but I
33:05 everybody realized that this great doctor machine , like crew scheduling, is a
33:14 task and in particular, when you that weather, so you have to
33:20 things put unquote on the fight to your plans. And both have to
33:29 planes and crews and restrictions on how hours people can work its September planes
33:35 can fly. And so whenever the , with some bad weather hits part
33:41 Egyptian doing, finding ahead for part it, you really have to do
33:46 really well. Things are happening. . So this is another kind of
33:53 that put in from a company I to work for. All right,
34:02 , so the point of this side best not only that, engineering designs
34:07 be extensive. Ah, this particular on that I work for a
34:17 We're actually one on one of his . Waas to be able to do
34:25 of a kind things very competitively. one thing if you Bill Carson,
34:30 build hundreds of thousands on them, if the first you are didn't get
34:37 right to be the Brother Fife and and maybe the second or third prototype
34:43 it and eventually to get things When you do one of a kind
34:51 , you can't do prototypes. Everything used to call it that is
34:56 of a kind, so we have get it right the first time.
35:00 these are not in time. Simple , because you need todo both fluid
35:06 and structural mechanics and, um, field my Merrifield electrostatic field competitions for
35:17 very fine geometries as well as short so fluid and damage this traditional application
35:28 HPC methodology. And this just gives somebody l the scale of things so
35:38 allow you that works with back of interested in HPC problem. There was
35:44 what's kind of the next challenge in of their hi scale, which is
35:50 can computing. But this is kind showing that the industry is It's not
35:55 academics that finding a good challenge for in science, but it's something actually
36:02 used, I And then I'll come to that in industry to for product
36:09 and in particularly the airspace. In day, coming back to train this
36:15 another one and not freight trains. doesn't work. All that move that
36:20 . But this is an example of speed rail nut, um, exist
36:25 many parts of the globe. The . S. Is a late
36:27 John Speed rain. But European nations someone human already been using these
36:37 And there is, when we think it again is being old technology.
36:43 it turns out that it's by no simple, too. Design trains and
36:50 that, you know up to in case, over 300 miles an hour
36:56 500 kilometers an hour. Plus, don't think, um, except for
37:02 rail based trains, something up to 400 kilometers an hour is still what's
37:10 and use the extremes on the Plus, I don't think is actually
37:17 use today, but the point of thing that they tried to me I
37:22 to make Is that coming back to wheels designing wheels There are in wheels
37:31 the tracks for these trains is by Simple, because you have to figure
37:38 that that the contact with the Raelians also in. That's in this case
37:44 there is a curb or a bend the track and the other part,
37:48 have to also make sure that because push a lot of air around these
37:53 , when they do at high speed you don't want when they meet another
38:00 , they won't go through a You don't want the pushchair to make
38:05 train derails, so you have to without the fluids, innovations around it
38:09 the mechanical simulations. Do you understand these things are actually things that you
38:17 have heard of? But otherwise you don't think of that Procter and
38:23 which is a consumer good, company that I think you say most
38:31 their products cost less than a and they use supercomputers on the high
38:39 computing technology in order to design things Pringles. So they're trying to figure
38:46 to shake the Pringles, so passion quick class, possibly in through
38:50 They also use s busy and so computers to decide things like diapers.
39:00 , some signs applications, and some on him. Biology, biochemistry.
39:05 this is morning biomedical imaging. that is trying to do things down
39:12 the Makram Ali to level well, even stand level. But to do
39:20 whole range basically from things that are the micrometeorite range or less to hold
39:28 organs that are again it kind off meter. And so several orders of
39:35 in terms of scales that are being in from sir geometry. And this
39:40 , in terms of time scales things down from two seconds up to,
39:46 given minutes or hours, depending on to do. Oh, and here's
39:51 example for people and used to work in the mid center. That electron
39:56 cop my crossed off the imaging. so today, what they managed to
40:03 is in the order on On that means I can result, comes
40:07 to about 10 to the minus 10 . And they do that from this
40:13 off that looks like just noise in upper right hand corner of what you
40:18 see in this light with the Jell rings and the yellow rings Is eso
40:26 underlying thing before the any other ring is what you get out of the
40:33 . And then, in this it waas early on them. It's
40:38 hand mark, where the actual iwas an image of a Mac Ramallah
40:44 Nowadays there was more automatic detection and eventually you get to what you see
40:50 the lower right hand corner, which kind of a rendering off the molecule
40:57 this case being imaged. So the why this is also good that busy
41:04 is that the first reflection Maybe that said, Well, this is just
41:11 , So why not improve this signal noise on? In order to do
41:16 , you need to put more energy , um, the's observations of the
41:25 . So this is not his. again electron beams, not X race
41:31 X rays. But X rays price thing. And again, the more
41:35 you put in Ah, the more you are to damage the specimens,
41:39 can no longer observed what? So ? So this is essentially what you
41:44 is to be able to extract useful out of a snow see images as
41:50 . So that means both are terrific computational challenges to make use for the
41:56 off. What this mostly knows on and other things more at the macro
42:02 . Trying to find out, you , a simulation of reading, for
42:06 , on this is my father that said in terms of the range many
42:13 , for in the full scope of science applications, things that were I
42:19 becoming aware of today with hurricanes are court approaching Houston is whether in climate
42:28 . And, um, this is both for, ah, human lifes
42:36 and also cost. But you can in order to you at least get
42:42 out of harm's way talking, able predict since, and this is on
42:47 , I can a big footprint So in that case, some we
42:53 what happened. I think people in numbers of people, um, tornadoes
43:00 a different kind of things. They a very that relatively were very for
43:05 footprint. Certain respect to compare two , um on that means you cannot
43:13 go through the type of evacuation if to do for hurricanes, so predicting
43:18 well the pathway or NATO's is another that is important for men.
43:26 situation the building climate. You can imagine that that support tears one
43:31 actually is Houston. Nothing I worked . People in the Earth and Atmospheric
43:36 Department provide monitoring Ah, air quality Houston. That's also something that require
43:46 Do you want to do it really ? It's a combination of prediction and
43:52 , the time monitoring or what to in order to not unnecessarily, which
43:58 happened in Houston shutdown school because are make sure that students they endorse
44:05 the air is unhealthy to the point you should not be outdoors. So
44:11 exercises, and you don't necessarily want shut down the whole state. But
44:17 being fairly detail as to where things bad on. This is just a
44:23 and more of that. Another thing is also, I guess, good
44:29 of us this time with wildfires being serious in cornea. So this is
44:40 an event. Ah, yes, by now. A long time
44:46 20 years ago in those Alamos there a gym. Trier and one of
44:51 comments that diamond that those elements unguarded a few years that you know what
44:57 similar it, what could happen in situation. But he didn't get an
45:04 of it, but it was very fire that hit and of what's in
45:13 picture toward the right. There's, fact, that happened to get out
45:17 at the time. So the smoke from this fire I was reaching even
45:25 10,000 ft to them is a So airplanes for flying around the smoke
45:33 because there's ashes and ashes in the is not a good thing. So
45:39 had to get rerouted even for flights , um, our tiny house did
45:47 ? Other things were lots of It's happened with astronomy, and,
45:53 you can see here, that terabytes good day in most recent type of
46:05 that are using both are picking telescopes radio telescopes like this one here.
46:11 are the most recent buns you That's the same with a low father
46:17 is, low frequency on 10 a that's an over 100 terabytes for date
46:26 . That's the continuously monitoring stuff I remember one defender, one of yours
46:33 physics this year. But the large colliders was one of the early scientific
46:41 for very large. Bigger, coming in. And it's in the
46:46 of several petabytes or tens of petabytes year now. So what's the benefit
46:52 this thing? You know, coming back to the sector that is
47:00 production. If one can actually improve care by a present, that's a
47:08 $100 billion a year and pay Here's another one from General Electric's of
47:14 1% improvement because gives across there products that the self, the customers we're
47:23 results in about $300 billion in There's a few more things that bestest
47:30 that there is potentially very good payoff using is busy in a number of
47:43 . So is what concrete that was in a study by now a few
47:47 old. This still gives you an all the value off investing unit,
47:56 , technology and just a little bit in terms of market segments that it's
48:03 all that much of the marginals Oh, so it's, uh
48:11 So when I was coming back a it to the scale much. But
48:15 I tried to emphasize in the examples shoulder in terms of applications waas that
48:21 scale aspects in terms of they have sizes more so than in actual computational
48:28 . But here's a little bit more that about one of the things that
48:34 talked about by the bit, which autonomous vehicles it's often referred to in
48:44 off its computational requirements as a supercomputer Well, so now to get to
48:53 come back to this game here. quick through here quickly. So So
48:59 is a picture of front off. of the supercomputers have supported the door
49:04 national lack, um, a few back Country. Now it's not the
49:08 recent one and the footprint Texas about tennis courts for the computer itself,
49:16 it takes another tennis court for the and cooling equipment to be able to
49:20 this computer and the cost. It's the hundreds of 200 million doors,
49:30 the most recent contracts ever given for excess care computers for $500 million.
49:40 , the energy consumption for this thing the equivalent of about 10,000 households.
49:50 it's fine omens, a small amount energy and the costs. And if
49:55 want to do, is green, it takes to set about 50,000.
50:03 , they credibly unsorted and listing is my view here, but,
50:09 and it's just the cost off the . So to try to cram left
50:15 a car is obviously something needs to . So that means reduction in size
50:26 , and cost needs to be scared what is a marinating. So it's
50:30 just as probably everybody knows, or so that has been around for a
50:36 time and have allowed for exponential improvement , um, first performance.
50:46 it's not going to cut it in next few years. So these innovations
50:52 happen both in algorithms, in computer , er and, uh,
50:59 essentially and have to. Not So where is another example that AI
51:07 now? I hope thing. And this was something published that's a year
51:14 two ago. A. Some took trouble when it came to natural language
51:22 and using machine learning for natural language in as you may know training tends
51:30 be the most potentially intensive compared to inference park when you use your A
51:37 learning all this to make predictions or that was currently does happening so that
51:43 was kind of interesting. Teoh give a little bit the perspective on this
51:48 . So what's in the left column the slide is if you machine learning
51:53 is that people use for a natural processing. Then there's a common for
51:59 may not mean the peace and the one hundreds they are, and these
52:05 to use and the teeth two years will chronic for machine learning. And
52:10 talk about that and after day, at some later point in the
52:15 But I want to pay attention to the Energy column that okay, wh
52:22 column. And so you want to on, uh, dioxide emission associated
52:31 the energy consumed was just a A rates. Um, come So they
52:40 some perspective on this number is what put under the table is what,
52:49 , the energy consumption is for typical household. So you can see if
52:55 use this neuron architect. The search The most expensive one. It's about
53:03 equivalent 16 your resumes just to do training search. And if you're can
53:14 look at the Carmen dioxide emissions that's prevalent. Teoh 700,000 miles off,
53:26 a typical of U. S So if it's just one month and
53:31 next one is kind of an um, but you can look at
53:36 sole. This is again for machine , not just the language processing,
53:40 it has a different set, the and it, says the actress.
53:45 get with the smiles today and then kind of have target schooling, former
53:52 terms, all basically improving the So when it comes to the image
53:58 applications are basically getting down to improve accuracy from one average 11.5% 5% and
54:06 can see this is, you look, in this case for competition
54:11 , they grow by my fable has magnitude and there, through Collins fallen
54:17 an exponential, and that's a little ah, you to the uncertainty about
54:24 the model section It's Caliphate soprano me scaling Uh, or if it's the
54:29 presidential skating's. I'm not going to into details, but just to give
54:34 some idea off his challenge, there's improve things in terms of both competition
54:42 and increasingly, also the environment. , this is a just example of
54:51 data in terms of where, how data on the race and the growth
54:55 and data the next year Slice gives some more concrete examples. No
55:01 In this case, the first concepts on a hallmark. Everybody knows
55:07 That's a few years back on I'm sure it's more today the processed
55:13 2.5 petabytes of data for our and about one million customers, their house
55:21 other things in terms of credit cards just for giving the devil scale
55:31 um, on such statistics searches We'll come back to that later in
55:39 examples and class. So all the companies Google that supports searching the Internet
55:46 Google and Microsoft and Alibaba and I . They designed their hard for that
55:57 order to keep up with the demand managing the power consumptions that they need
56:02 make it more efficient than their standard you, but to be able to
56:07 their job. So I was a . Nagle quickly about the bunch of
56:14 applications and, uh, trying to the lift More specific lists of what
56:20 I mean with as busy and what said. See nothing of what I
56:25 with it. But then it I would like to stress the resource
56:30 aspect on the technology that is used achieve that. So a lot of
56:38 , as you can gather from these I have given esta, is going
56:43 in the applications. And, unfortunately, Stanley compilers are Madras.
56:54 successful in doing that. So typically requires the programmer to help out and
57:01 colds and express things in a way the jewels sank and the details in
57:08 of mapping things onto the particular Portland on. It's also questions using algorithms
57:17 again, from my examples, managing Izabal concurrency at the Lord Scale.
57:27 is increasingly important. And that's not in terms of science and engineering.
57:32 the Internet companies that sold with visit with months of tens of thousands of
57:39 for a minute Now, the thing , as you shall not been addressed
57:45 often still isn't in many off cause on Esposito. Probably computing is the
57:52 consumption. But that husband, I say the focus on both the,
58:01 , silicon industry left until I am , um, and media for over
58:10 decade. And as well platform members their legs be, um, and
58:20 one tends to do business with Um, but also, in terms
58:25 the end users, the energy consumption pretty much the number. Not one
58:32 these days are design at any level this chain. So I will include
58:39 number of things on and through the , trying to get him aware of
58:46 consumption and even be able to observe consumption off your coats. This is
58:54 , and I saw, um, Nuclear probably see it on the
59:00 This dislike this in this case not look at the scale off application demands
59:09 terms off computation or data sets, in terms off the size of computer
59:19 . Of course, the whole crime what it stands for If the 500
59:26 choir fools systems are, there's a published twice a year. And,
59:34 course these are relatively large, clean . So you can say Well,
59:40 I got the years and one of when the company several work for made
59:45 don't have it, either that, , Internet companies do on the computer
59:51 do, and then as well as labs and many other interests. So
59:58 this line is essentially just the first . I wanted you to pay attention
60:04 Seoul. If you look at their powerful systems. No, no,
60:11 should say There are also many systems never get uh, published, some
60:20 speak, because there are operated by for defense purposes and they don't tell
60:27 they have. But anyway, you see in this case or one of
60:31 systems 10 million courts and course basically , can you? That's the processing
60:40 . The corn is effectively the complete . It's just Pakistan on together with
60:48 processes in a piece of silicon. there is best 10 million execution streams
60:54 more in this system. So when talk about your coast being able to
61:00 , of course, you may not that. But even in the smallest
61:04 on the system. So if you at the bottom 10 system on this
61:09 list, they're still tens. If 10,000 pores and those reluctantly small
61:21 Yes, and this society was best point that it's not just governments or
61:29 that has these kind of systems. ends up on the stock.
61:35 This is, in fact, more half of the systems on the
61:39 Our systems in industry and these are companies, aerospace companies, automotive
61:46 energy companies. So thank you have is just a living example of
61:55 uh, snapshot taking from one of National Department of Energy of getting sentence
62:04 a few years back on. This that some of the coals here didn't
62:09 so. Millions, of course. the use tense off thousands, of
62:13 . And the next side similar very . So the third line from the
62:19 tells you how many course there Ah, colds are using and best
62:26 the frequency of colds. Um, uses, uh, course in a
62:32 race. So, as you can , there are cold colds that runs
62:39 this particular computing centre nurse that used than under 1000 course. So you
62:46 to have again applications to make sure scale it up all right, a
62:53 bit more So now. So that about kind of the scale. Also
62:57 computing systems and came ability. So I was a little bit about performance
63:05 I said, Hi. Energy efficiency also important. But high performers means
63:11 efficiency, high resource utilization because at level of ah, many systems
63:23 you cannot have inefficient codes and make for it by buying a bigger
63:31 I again and you met many Um may not use even 1% of
63:45 systems that are basically your waist 99% your money and energy, and that's
63:52 not acceptable. It's never been I would say, in mobile
63:59 But before energy become a serious issue stationary systems it was often ignored.
64:08 it's no longer in north by So here's why. Here is a
64:14 from condoning that this chief technology officer NVIDIA not missing one are you can
64:23 anyway, so you know often gets good. Um, cancer may good
64:29 um, build. Ali used to a profession. I think just alone
64:36 a professorship at Stanford. He was to that professor than my d and
64:41 cramp your student that contact. So is a very prolific. So here's
64:47 deal. You know, in terms energy, this is not so much
64:52 Internet companies and many companies are and these large state of centrists they
65:01 often tends off mega. What and usual rule of thumb is and making
65:12 , um, data center. Over year, the energy costs about the
65:20 dollars and some million dollar per So in case you use a tense
65:25 that, that's tens of millions of you spend on utility bills. And
65:32 much money, even for Internet It's so you're so what they do
65:39 from Google Facebook. And if you seen one of these data centrist,
65:45 is kind of a little bit to collage of pictures. So the point
65:49 it is not only to see lots computers, but in fact, the
65:54 empire infrastructure requires lots of what you is a lot of plumbing, and
66:01 the pool So this is what serious centrist looks like. It's not just
66:09 , and here's the other part of about the energy. So it's
66:15 I was in in most parts of globe more than a decade ago that
66:23 energy consumption for power and cooling costs than the computer system itself softer and
66:33 . So the total cost of ownership the lifetime of the system is dominated
66:39 energy consumption. Now, by the on the system, you get a
66:48 , there's a couple of quick slice go through in terms. So why
66:51 consumption has become important on this is side. It was done most of
66:58 pieces not for Custer's, but Essentially, it's the same engine started
67:05 used, but there used to spill server for his PC or customer or
67:10 not, and it's sort of more less follows. More is low in
67:14 off Harvey Energy consumption has improved over years. Uh, this is not
67:22 true for this larger scale systems are there. This thing for the most
67:28 top 500 list that was published in back to the initial part so it's
67:35 quite as good on evolution as for . But even, um, if
67:44 looked at the stop from in the my different angle. So this is
67:50 growth rate of capabilities of these systems the years. Of those three lines
67:55 , dark rolling at the bottom is the number 500 system that the least
68:03 system on the list on the triangles the brownish or whatever in sort of
68:09 middle range is second probabilities off the powerful system on the list. And
68:17 the green dotted line on the top basically the aggregate of all systems on
68:23 risk and the improvements over the These exponential improvement, but there was
68:31 doubling about every 12 13 months for long time. It's no longer
68:36 but there waas until 57 years That means more snow gave you something
68:43 18 20 months, doubling, but capabilities doubled more a lot faster.
68:49 that's why car consumption again became a for data centrist because they are the
68:58 between improvements for more slow and system . I was significant. So what's
69:05 20% a year and that's an So that's why many data centers run
69:10 of power and have Teoh rebuild their pretty much for every generation computing system
69:18 got, um, a little bit a couple of more minutes, and
69:25 , uh, hopefully I will try finish up, and this is just
69:30 that will talk about it. So did this thing happen? Well,
69:35 happens because of the underlying technology that being used, that is for short
69:41 , most complimentary in metal oxide semiconductors the way things work for a long
69:50 . Waas The power is proportional today off the hostage by which the circuit
69:57 silicon operates and the cop frequency being . So that's what's no less the
70:02 low. And that's in a useful to remember on for a long time
70:08 worked in. Designers favor, so speak, that can increase clock
70:13 Ah, the performance. Increased features smaller, so you got your facts
70:22 capability at the less cost for less for a very long time it
70:31 but it's not true anymore, and why we now have most. According
70:35 talk about that and they? They're some. So what happened? This
70:40 a little bit this picture from um the if you look at their
70:50 to red, Not too deep red . That shows you basically useful work
70:57 walked for us. Several generations of processors. So for many years,
71:06 image efficiency off the chips actually went with next generation shipped up there.
71:15 . And I'm fortunate. So, , off the chip companies Napoli
71:23 um about then 20 years ago was aim the and intel in for one
71:30 the last ones, actually to I that this is not working. So
71:38 , in fact, had to round abandon one of their product generations because
71:42 couldn't call their chips. It was the readies to be marketed and
71:48 so they have to go back to drawing board. The other one's
71:52 Minnie Andy actually started to produce multiple ahead of intel. So what they
71:58 , um, we got montecor and also ended up taking out some of
72:04 features and their processors to improve the feet efficiency. And, um,
72:14 sign is very quick to say part the things over the generations or the
72:19 wants to get bigger and bigger markets the Babel to sell more and more
72:24 of the same chip that you added . And this Silas consent the people
72:31 one of the national arm that I be before the nurse. Well,
72:36 they discovered that they only needed about off the 300 instructions at the time
72:43 most in an Intel processor for their to do a good job so you
72:48 throw ups based with 75% of instructions on it was still just And this
72:58 shows the range off processors. So science is coming back toe,
73:05 this so improving energy efficiency is that have to then, as an individual
73:11 , you cannot do much about But you can select what type of
73:15 you're using for your Marghera and it this tells you the difference in capabilities
73:26 so performance and former off this case instructions per second he construction per second
73:32 the power and the level of special . So it ah, basically the
73:45 nine. I think I had better excited best assess that if you do
73:53 , you can get 3 to 4 of magnitude. More performance for
74:02 So you get more stuff from the and we'll come talk about that much
74:09 . So General Microprocessor is something wrong them, but they're not particularly
74:16 Um, and many ones are, know, resorted to keep use that
74:21 kind of specialist processors. They're not flexible. Let's make microprocessors. There
74:28 have been more energy efficiency efficient than processes, but not by much because
74:36 are still very power hungry. So there are other things like digital single
74:41 and more dedicated process is known as six ritual thought so part of what
74:49 ? The head Virginia's knows that on is again due to the fact that
74:55 needs to be more energy efficient. I just I will stop at this
75:03 or some few more slice in the deck. But as no pipeline is
75:12 popular programming language, and there's nothing with fighter on its very good
75:18 Most times it's not very efficient. for me, not Jews in fighting
75:23 the classical, other things. So was a study done by A
75:29 I t. In a couple incarnations the last nothing more or less
75:34 But that was published recently, which guess I forgot to put on the
75:38 . That would be that it This particular table came from on article
75:44 science and June this year, but based in shows that there is about
75:52 orders of magnitude difference between a naive and cold and a highly often my
76:02 . So in the course of this , I hope to be able to
76:09 you insights and ideas on some Go on how you can get from
76:15 Piketon performance to There are upper right corner, and you may see several
76:22 of magnitude improvement and the running time your coat. And with that also
76:30 reduction in energy footprint. It's a . So I will stop project since
76:36 kindness up and be happy to take . We should have left one more
76:44 for questions. Um, China's just a painted picture off the back
76:52 the class on the emphasis on software and the fact to get the deficiency
77:02 need to understand the underlying platform. , We'll surf how? This is
77:19 . You are here. So I a question regarding red domes. Are
77:25 Ah, in the schedule you have that you'll be giving momentum and there
77:30 be a due date for the metal the victim would not have ah dining
77:35 something. We have to submit You know, I ve crucial,
77:40 ? Get the weight to do the . So it's not something that you
77:45 , you know, an hour an and a half. One class time
77:48 do. But it's on the flavor take Hold him in terms old typical
77:54 you need to do Some need to through lecture slides and potential hand us
78:00 you have been given and potentially also stuff up on your own on the
78:06 to answer the questions in the Okay, Okay. And don't final
78:14 would be the final ending of the . That's correct. That's all you
78:19 dio that's going to find a project . D should be ideally trying to
78:26 it started enough that you have about weeks result to do the project.
78:31 , do the projects and then you the project report on during the time
78:37 for the final exam, you will the presentation of your project.
78:42 so not just to me and Josh, But so everybody is expected
78:47 be there and learn from each other terms of what the projects were and
78:53 and successes. Okay. Okay. you. Okay. Dr.
79:01 Uh, this is Ivan Rodriguez speaking the final project. Um, it's
79:06 , already. That's pretty open Um, if we work in
79:10 can we do something in conjunction with company that we work for assed
79:14 That's related to each piece there? . And as long as you're allowed
79:19 talk about it so they don't you . Okay. No. Yeah.
79:26 . So intent to encourage not um, students that comes from
79:32 but that works. Many of them in various research groups of the
79:38 So if they have a problem that would work on any way and for
79:44 reasons, we can just stress the aspect. So one students to do
79:50 that they are interested in, rather me telling them what to do.
79:55 . Current thank you, but I also after that, if somebody doesn't
80:04 a strong sense for what they want do, then I will certainly
80:09 um, potential projects and have a of all students have done in past
80:16 that maybe the way our stimulating thoughts interested, what you might like to
80:23 so. But again, I welcome own suggestions. So the former it
80:29 that you will write up the paragraph two. But what you plan to
80:34 , how you, um what a in our your plan to do it
80:41 also the resources needed. So, , I need something that is not
80:48 kind of provided by the course. only happen to try to help figure
80:53 how to get resources you may um, or otherwise will need to
80:58 , Well, this is not feasible do what we have available, so
81:02 need to figure out something different. I tried to guard about it not
81:09 within the scope on the course. are too ambitious and is not
81:16 for the time allowed it for the . Andi, not so well enough
81:25 you will get some insights that you get through the assignments. So tried
81:32 make sure that this little write up you do before you start on the
81:36 . That is a balance between Resource is and this scope from the
81:53 and the other questions currents, if some thank you for is another question
82:09 the chat from silently she's asking can used for Quincy l for the assignments
82:15 the project. Can we use Open cm open? Cmv C l
82:28 , yeah. Again, if it on. So for the See
82:39 Ah, I presume that thinking well for the you programming exercise.
82:50 and since I discouraged the use of because it's a proprietary language and I
82:58 like Teoh, uh, teach proprietor . So opens here is an open
83:05 language. So I think for the we do, it may work.
83:15 and it depends whether open seal is on opens Soul. It's unfortunate ConStor
83:24 It's very tricky. So I believe video and principal supports open C
83:35 But I'm not quite sure how good implementation isn't. What extent that this
83:40 computing center actually do support open CEO they only have indeed edge abuse Their
83:49 JP you manufactures out there. and if you have access to something
83:56 has good support for it open C and you on your own,
84:03 then I think that the okay to it. Um, but it is
84:08 big company. It's a big Unfortunately, open a sissy as becoming
84:15 open. It wasn't when it It was a proprietary thing again,
84:21 , open. And he has been since the beginning on 34 years
84:30 This started to also support the programming accelerators. So the standard itself addresses
84:41 Virginia's notes. But the compilers use open MP on the platforms that we
84:51 available may not necessarily support open and for GPS. So it keeps
85:01 Um, but it ends up being of literally a mess in terms off
85:12 is available at various ice. And depends on the vendors so much so
85:22 think that coming in the cities, to figure out how to move away
85:29 it, But as always, it's tends to start up. We're trying
85:36 do vendor lock in until it becomes and then eventually came up from making
85:45 source so broke up ahead Virginia Snows still kind of a mess.
86:00 we'll talk more about it. I'm get to talk about you to use
86:04 town. So it's not the you can use it if it
86:10 I'm not the Post. I just open CEO in the class. I
86:24 see any questions. I would just remind everyone to get their accounts and
86:28 and send your user names to assess his friends on. Did you
86:36 you know, additional question that pops your mind after the class sender's
86:41 and I feel want personal interaction. know, saying the moment, configure
86:49 , time sooner or whatever. One one sessions. Okay, it's
87:03 No more questions than no stop today and first on the block the Giants
87:13 stop the recording properly
-
+