YT YouTube2Text

YouTube2Text → Transcripts → [CS61C FA20] Lecture 37.2 - Data Centers, Cloud Computing (WSC): Warehouse Scale Computers

[CS61C FA20] Lecture 37.2 - Data Centers, Cloud Computing (WSC): Warehouse Scale Computers — Transcript

by CS 61C Departmental · 5,502 words · 881 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00and welcome back now let's finally see
0:03it
0:03a warehouse scale computer how does it
0:05even work
0:07so why cloud computing now let's talk
0:08about why this didn't happen 10 years
0:10ago or 20 years ago
0:12um these extremely large data centers 10
0:15000 commodity pcs
0:16um more users built out of the demand um
0:21in the early days so let's go back 20
0:23years there were
0:24every company used to have their own it
0:27department they probably had some
0:28servers in their basement
0:30and they ran their they basically rolled
0:31their own everybody every kind of
0:33reasonable size company was rolling
0:34their own
0:35and then as you had big companies the
0:37google's the amazons of the world
0:40i guess apple and somewhere in there
0:42facebook is in there as well
0:43they realized they needed a much larger
0:45scale than just a basement you know set
0:47of computers that they'd have
0:48and that large scale allowed them to
0:50think about well as we're growing if
0:52we're paying this much money for it
0:53what how should we build these well they
0:55started to build them out of
0:56commodity pcs they re i mentioned before
0:59they realized that build not having the
1:00highest end gaming machine and putting
1:02you know multiples of those in there but
1:04having the cheapest best value per
1:06dollar
1:06really it's about commodity pcs um and
1:09they just decided
1:10there's a higher higher failure rate for
1:12these higher for those machines but
1:14we'll live with that we'll work with
1:16that um it's definitely better so the
1:18economy scale was
1:19um it's it's it's actually five or seven
1:22times cheaper
1:23than provisioning a medium-sized
1:24facility per per machine
1:26per instance they also had more
1:29pervasive broadband internet
1:31so they could actually you know access
1:33these things uh
1:34efficiently and again commoditization of
1:36hardware software
1:38to help help with that it was certainly
1:39standardization of software stacks
1:42so i looked up recently what the
1:44instance prices were
1:46if you want if you did the reading then
1:47you'll see that they talked about a
1:49particular example in the reading the
1:50warehouse scale computing
1:52and this is amazon web services aws and
1:55this is right here this guy this
1:58particular one here
2:00is the sim is the example that's closest
2:02to the example they talked about in that
2:04reading
2:05which is about 33 cents an hour
2:08or three to three three hours per dollar
2:12in terms of the ratio of the smallest
2:14you can buy a very minimally provisioned
2:16machine
2:16a very military machine you know the two
2:18gibby bytes of ram
2:20two cores very small very small cheap
2:22two cents an hour
2:24remarkable uh but if you want a little
2:26bit more you know a little bit more
2:27power
2:28you've got uh you know eight virtual
2:30cores 32gb of ram
2:32i appreciate that but again three you
2:34know three hours per
2:36dollar so it starts to it starts to add
2:37up if you leave it if you leave it on
2:39and forget forget it's running and you
2:41know i can probably leave this on
2:42at this price is 16 times more expensive
2:44i could leave this guy on
2:46uh for a while and not worry about it
2:47too much but the one below it you really
2:49want to watch out for that
2:51um and what's incredible is and so
2:54there's a whole you can you know there's
2:55tons of knobs you can choose i want to
2:57have more memory i want to have higher
2:59high cpu and so you can
3:00decide what you want they do give you
3:03what's what's called
3:04ec2 or compute integer unit this is a
3:06number kind of a spec
3:07a specification tells you what you're
3:09buying um and so you can really say well
3:11i need to have this much power
3:12but a high memory system well not
3:14actually this power because of this this
3:15is a cpu
3:16it's really a cpu you look at your data
3:19you look at your data and you say
3:21what what is this computation really
3:23needing is it needing a lot of memory is
3:25it not a lot of compute
3:27i don't know a lot of cpu cycles but
3:28it's a lot of memory load the whole
3:29matrix in to do a couple of searches but
3:31i don't have to go to disk
3:32so i want a high memory machine or maybe
3:34it's not a lot of memory needed but a
3:36lot of cpu but i'm going to be crunching
3:37on
3:37a couple small data sets so you're
3:40playing with where you want high cpu
3:41high memory or just a standard setup
3:43that's a little bigger than that so you
3:44decide what you want in terms of that
3:46but there's a whole scale there what's
3:49amazing to me is that at these low rates
3:50it feels like low rates to me to have
3:52just a machine somebody's going to keep
3:53a machine
3:53and they'll keep all the worry about
3:55making machine up failures all those
3:57things they'll take care of it i just
3:58have to rent the machine that's a pretty
3:59good deal
4:00i appreciate that just for compute or
4:02storage too they also have
4:04elastic block storage um and they have a
4:0710 cents per gigabit
4:09uh this is gigabit per month or
4:12for a hard drive or a spinning drive
4:14it's four uh four
4:15four cents um gigabit take a bite sorry
4:18gigabyte gigabyte per month
4:20gigabyte month i said gigabyte month so
4:22if you want more gigabytes or more
4:24months
4:24it's a gigabyte month remember that it's
4:26not getting by per month it's gigabyte
4:27month
4:28also important you can also get any of
4:29these machines with an attached ssd if
4:31you need to have some local stuff that
4:32you play with and you can choose
4:34and explore your space with that
4:37so again here's the idea you've got this
4:39massive data center
4:41100 000 to 10 000 servers all in one
4:43building
4:44um emphasizing cost efficiency but you
4:47really need to pay attention to power so
4:49cooling keeping these guys cool is
4:51really where a lot of your engineering
4:53goes
4:53and a lot of your thought goes a lot of
4:55your square footage goes is to thinking
4:56about keeping those things cool
4:58um it's relatively homogeneous so yes
5:01okay that that was version one of the
5:03commodity version you know upgraded them
5:04so now as i upgrade maybe some of them
5:05are still version one some other version
5:07two but mostly they're homo you know
5:09homogeneous in terms of hardware and
5:10software you offer
5:12incredible amounts of software as a
5:15virtual platform
5:16uh social networking video sharing you
5:18can allow companies to buy
5:19time on the on your rack and this is
5:21very exciting for many companies
5:24very high availability we call it nine
5:26five five nines of availability so
5:28less than an hour of downtime per year
5:30which is credible and in fact sometimes
5:33you know you can look up on some of the
5:34news articles where you see oh uh if
5:36amazon goes
5:37goes down you know some particular
5:38amazon warehouse slows down what will
5:40happen
5:40is many companies will go down with them
5:43why was this site offline well because
5:45it's
5:46served on that particular thing you
5:47remember that there was a flood on the
5:48eastern seaboard that took a lot took
5:50away a lot of services and the services
5:52go down with it so it's not just
5:53amazon's going to suffer that they've
5:54got a ton of redundancy across that you
5:55probably won't notice on the amazon side
5:57but you'll notice on the smaller compute
5:58side they might have had they might not
6:00have been paying for that redundancy in
6:01their systems
6:02so sometimes services will go down
6:04because they're running on the aws
6:06local system which itself went offline
6:08who knows
6:09anyway two researchers said warehouse
6:11scale computers are no less worthy of
6:13the expertise of computer system
6:14architects than any other class of
6:15machines
6:16you could deal with a really small micro
6:18level they're kind of the the smart dust
6:20level
6:20all the way up to a particular cpu we
6:23spent a lot of time talking about
6:24building it building
6:24faster chips to building a faster
6:27computer with a lot of chips in it to
6:29building
6:29now a warehouse of lots of those
6:31computers so that's a it's a fun it's a
6:33fun space it's at the macro scale
6:35we're dealing with here now i give you
6:38a really fat check 10 million dollars or
6:41more and i say build one of these
6:42warehouses for you
6:43what are your design goals what are you
6:45thinking of when you're building these
6:46this warehouse scale system
6:48a ton of parallelism just incredible
6:50number parallelism largely number
6:52independent data sets
6:53independent processing we call this
6:54again data data level process data level
6:56data level parallelism what are the
6:59issues of
7:00scale and what are the opportunities um
7:03well there are not many of them if you
7:06may say well go look around the world
7:07for other
7:08warehouse scale systems to model years
7:10after there aren't that many of them
7:12there's just a handful of these very
7:13wealthy companies building these systems
7:15um so it's not it's it's not like well
7:17there have been
7:18four billion in uh implementations of a
7:21particular
7:22chip let's look at the four billion and
7:24first okay that's i looked
7:26what did they do wrong what did they do
7:27right this is hard it's hard to if
7:29there's only four or five of these that
7:30has ever occurred or four or five
7:31different models they've ever occurred i
7:33mean there's certainly many more of this
7:34that totally occurred but in terms of
7:35how many different models you've got a
7:36facebook model you've got amazon's mod i
7:38got google's model apple's mob
7:39not many it starts to sh how many more
7:42there's not that many more of these big
7:44big companies running these big
7:45warehouse scale systems
7:46um you can certainly get price breaks
7:49uh from purchases of commodity thinking
7:51look i'm about to buy a million of your
7:52computers
7:53how low can you go because if you don't
7:54go there i'm going to this company so
7:55you can keep playing against each other
7:57to get some of those numbers down
7:58um and you're going to have a high
7:59number of component failures so you have
8:01to have somebody and by the way if
8:02you've seen these i'm going to show you
8:03some videos at the end of this
8:04there are people videos of people who
8:05are replacing drives because
8:07there's a job who's this there's a job
8:09whose title is you're the drive
8:11replacement person
8:12now you could be worse you could you
8:13know be shoveling something you know
8:15shuggling poop or something but
8:16the point is your job is just to skate
8:18around and they have
8:19rollerblades and they skate around and
8:22they bring the
8:22hard drive swap bubble up done hard
8:24drive swap into a raid system these raid
8:26systems allow for hard drive failures
8:27without data loss
8:28something normally if you don't think
8:29about a raid you have a single drive and
8:31your home computer
8:32once it fails you've lost data well not
8:34if you have redundancy there we're going
8:35to see this in a lecture a little
8:37bit later i can have enough redundancy
8:39that can have one or maybe even two
8:40fails depending on what level of raid
8:41you have
8:42so you can lose it drive replace the
8:43drive as long as you don't have past a
8:45certain limit
8:46then you won't have any failures so this
8:47whole system this whole particular
8:49configuration
8:50you could have a drive fail and because
8:52i have redundancy i don't lose the data
8:54that's amazing so
8:55you have somebody whose job is just to
8:56skateboard around
8:58rollerblade around and replace drives
9:00pretty incredible
9:01good job fun um here's the other key
9:04piece of this is
9:05the cost of equipment purchase the first
9:07you know the check you write is not a
9:09one-time check to build this system i
9:10built the system
9:11but now the cost of ownership is far
9:14greater than the cost of
9:15purchasing it initially you've got power
9:18you've got cooling you've got
9:19manpower all those things replacing hard
9:21drives all those ownership the cost of
9:23ownership is far greater than
9:25the cost of the initial purchase so it's
9:26not just well thank you for the check
9:28but i need to i'm gonna be needing a
9:29check uh every month to pay for
9:31all the overhead i have to put into the
9:32system here's the first picture
9:35here's google's horizon google's oregon
9:38warehouse scale computer
9:40and what a couple things you notice
9:43it's in the middle of nowhere first
9:45thing i'm seeing
9:46it's beautiful it's beautiful picture
9:48it's middle of nowhere
9:50and the reason for that privacy you've
9:53got
9:53you've got corporate sensitive data here
9:55who knows who what company is borrowing
9:57your system to put their stuff there
9:58they
9:58don't just have an easy location you
10:00want to be very hard to get there that's
10:01number one
10:02two land is cheap you're the middle of
10:05nowhere land is
10:06cheap so go buy if you have this much
10:07land go buy it where the cost of the
10:09land isn't very high where no one's
10:11gonna competing with you
10:11put this in the middle of the city
10:12that's the worst place to put it ever
10:13obviously
10:15it's in oregon what's oregon telling you
10:18it's cool in oregon
10:19it's not in you know the the heart of
10:22the hottest
10:23state in the union oregon's pretty cool
10:25place um
10:27so temperature is an important thing you
10:29don't have to you don't
10:30put this in a place where there's gonna
10:32be snow and you have to deal with
10:33keeping this warm blah blah blah there's
10:34a little bit of that as well but you
10:35don't want to
10:36deal with a lot of temperature
10:37variations and you're near a river
10:40you might not have thought about that
10:42what's that mean well maybe you can be
10:44very clever in terms of bringing some of
10:46that some of that water in
10:47and helping with your cooling system and
10:49maybe warming it out there so there's
10:51maybe a way to keep uh to work on saving
10:55saving um saving cost by using can be
10:58water controlled uh
11:00water controlled cooling thinking about
11:01that in fact here's the zooming here's a
11:04zoomed in version of that
11:05and then if you look down here if i zoom
11:07into that even more you know what that
11:08is
11:09that whole building in fact you can see
11:10that from a distance that whole building
11:12is the cooling center that is such a big
11:16part of this
11:17that keeping these servers which
11:18generates so much heat
11:20cool requires an entire building to just
11:22refrigerate them that system so
11:24we'll talk about that a little later but
11:26that's a very big consideration design
11:27consideration is making sure you
11:29you've handled cooling well inside what
11:32they've discovered
11:33is they build these containers you know
11:35the same shipping containers you see on
11:37the big boats taking things from to and
11:38from
11:39uh other countries on the on the oceans
11:41they decided to
11:42just and by the way we had one in soda
11:44hall we as we had a trial
11:46uh container in the backside of sort of
11:48hall um
11:49for a while this is we we used to have a
11:52sand volleyball pit then we had a
11:54warehouse scale
11:54we made one of these containers with
11:56with servers in them to practice to
11:58to play with them to explore that space
12:00um so we had one of these at sodahold
12:01for a while it was really fun
12:04so what you've got is they stacked them
12:05too high which is an interesting model
12:09they uh have
12:12access they put them in and had put them
12:14in they don't just put they put them in
12:16an angle it's almost like
12:16when you park a car in a parking lot and
12:19then inside a container is just an
12:20incredible incredible high density of
12:23servers
12:24um typically what you're seeing by the
12:27way
12:28is this the back they put them all
12:31this aisle this aisle is all the user
12:33facing parts so this is all the parts
12:34where you would
12:35maybe swap in a hard drive and do that
12:37and then the back side where the usually
12:38exhaust is and the wiring is
12:40that's on the far side so there's not a
12:41lot of wiring you're seeing here the
12:42wiring usually is on the back side of
12:44that so that's the nice thing
12:45also this could be sorry this could be
12:47the wiring aisle i'm looking at it here
12:49i think this is the wiring aisle see all
12:50those wires that come out of that there
12:52is another side so
12:53there are different this is the the back
12:54side out that's the front side it's the
12:56back side is the front side aisle you'll
12:57see that
12:58um so this is you have to have better
12:59you have to have easy flow you can't
13:01just have well i can't move it all in
13:02the wiring aisle you may have
13:04flow and move as you wire and you
13:05tighten them up you better be very
13:06organized with your wiring of it
13:08you have a lot of cable ties that keep
13:09these things otherwise it can be insane
13:11how many wires go around
13:13there'll be another aisle which is where
13:14you'd be swapping the hard drives out
13:16that'd be the front of these servers for
13:17there but that i think is the back side
13:18with all the lights i believe
13:22so what do you see inside that system
13:24you saw kind of a big picture of of one
13:26of the one of the aisles
13:28it starts with a server these are often
13:30called 1u servers they're about one and
13:32three quarter three
13:33they're very heavy by the way and very
13:34deep it's much deeper than you realize
13:36you think oh it's a computer it's a gray
13:38beige box on the side of you
13:39no not at all at all it looks like a
13:41pizza box
13:42um it's about one and three quarters
13:43inch wide night high
13:4519 inches uh wide and the size you're
13:48six feet
13:4919 inches wide and then 16 to 20 inches
13:51deep so it's really very heavy
13:53uh very flat uh inside that is a normal
13:56computer you've got an
13:57eight cores you know 16 gb big bites or
13:59more of dram and maybe four
14:01one terabytes or four four terabyte
14:02disks that's the first level so that's
14:03that little slicey guy
14:06then you've got a rack this is often
14:08seven feet high and you put 40 to 80
14:10servers
14:11in that rack um you've got a local area
14:13network of ten one to
14:15gibby bytes per second give me bits per
14:17second a switch in the middle
14:18and that's called the rack switch so
14:20right here is your switch okay
14:22placed locationally right in the middle
14:24of it so it's easy to get to and also
14:26kind of equidistant from all of them if
14:27you think about that space
14:30finally you've got an array and that
14:32array we often call a cluster
14:34that's 16 to 32 server racks in that
14:37and there's also then this would be a
14:38larger cluster switch these are the
14:40racks
14:41a larger and much more expensive cluster
14:43switch to be able to hand that
14:44typically 10 times faster but the cost
14:47is a hundred times because
14:48there's not there's fewer of these and
14:50the people say well you need it you got
14:51to pay it through the nose for that
14:53so it's a function typically n squared
14:54in terms of the cost for that
14:56switch versus the local switch here's
14:59again a picture of a server
15:01sra server there's a rack and there's an
15:04array
15:05and i believe this is the front side of
15:06that this is the user serviceable so you
15:08see
15:08the wiring is on the back side and this
15:10is the front side i put an f there it's
15:12the front side of this area
15:14and there's a lot more room here between
15:15them they want to be able to get between
15:16you see this little air that's that's
15:18loss of space
15:19so typically if you have the higher end
15:21systems they don't have that those those
15:22racks but this is kind of a smaller
15:24scale
15:24cluster what's inside one of those
15:28machines
15:29so that's the traditional way you build
15:31them when google did it
15:32google said why do we need the case
15:35let's build it ourselves they literally
15:36build much many of these things
15:37ourselves
15:38so here's a google computer if you can
15:40see what are you seeing here
15:42what am i seeing well there's your power
15:44supply
15:46uh there's your two cpus okay two double
15:48cpus there double fans on that
15:50here's our memory obviously here's our
15:53here's our hard drives
15:54um and here's some connection isn't that
15:57nothing else interesting
15:58here's a switch that here's a has a slot
16:00that isn't being used
16:01um is that there's another probably
16:04something there that's a that's a heat
16:05sink on that
16:06maybe some gpu possibly
16:10this this is right here this is the most
16:13important and interesting piece of it
16:15it's a battery so this is
16:18and you can tell us about it because it
16:20has just two leads to it right batteries
16:21typically don't have more than that
16:24this is their version of an
16:26uninterruptible power supply
16:28they've thought about having one big you
16:31know unintentional power supply
16:32basically says the power goes out
16:34how do we handle this how do you both um
16:36clean the power sometimes the power
16:38is not a perfect 60 hertz uh 120 volts
16:41peak to peak
16:42what you typically see alternating
16:44current you typically have
16:45um that someone has noise in the system
16:47that can be a problem for that
16:49and what happens when power goes out you
16:50don't want to lose your system you know
16:51let's say the power just bloop power
16:52went out for it for a second for five
16:54seconds for a minute for an hour
16:56for a day you want to be able to keep
16:57your computer up for as long as you can
16:59for the battery there's obviously a
17:00limit to that
17:00um and they played with the experiment
17:02with this at a hospital level
17:04you put a big uninterpretable power
17:06supply in the hospital basement
17:08that powers the whole hospital you don't
17:09have a ups at each floor at each table
17:12you don't do that you put a big one in
17:13the basement because it's a hospital
17:16google and other folks have tried that
17:18having a big ups
17:19and they realized it was more efficient
17:22to have a localized ups
17:24for each one there's something about
17:25scale and cost and
17:27and uh and heat and other elements that
17:29are involved with putting a big massive
17:30ups in the basement to handle the power
17:32for the whole system
17:33the power draw is incredible they
17:35decided it's actually more efficient
17:37to buy a localized ups battery supply
17:40here so that is the
17:42uninterruptible power supply in case
17:43power goes out they can still have that
17:45computer compute for a little while
17:46before the battery eventually dies
17:49so now let's take a step back and think
17:51about what performance is um
17:53what does it mean to say something is
17:55faster than something else so here's an
17:57example
17:572009 ferrari 599 gtb
18:00holds two passengers and 11.1 seconds
18:03for the quarter mile
18:03let's call it 10 seconds make it easy
18:06and this is
18:07a school bus 2009 type d school bus
18:1054 54 passengers quarter mile time say
18:13it's a minute okay
18:16when you say what's the performance of a
18:18system we have to ask it's almost like
18:20you tell the computer on star trek you
18:21say i want tea earl grey it'll say
18:23hot or cold two different flavors if i
18:26say i want to know this is the best
18:27performance they say well do you mean
18:28response time
18:29what do you mean latency sorry
18:32do you mean response time or latency
18:34which is the same thing or do you mean
18:35throughput
18:37so response time or latency means the
18:39time between start and completion of a
18:41single task
18:42i just need one thing i just need you to
18:43take this
18:45this raw metal and make it into a shape
18:47well i'm just
18:48i'm not doing a million i'm doing one of
18:49them what's my total latency between
18:51when i give you this
18:52until it's all done in this case what's
18:55the time between
18:56taking and moving a single person a
18:58quarter mile that would be the response
19:00time
19:01throughput says how many of these can
19:03you do over time so it's amount of work
19:05in a given time and for here we're
19:08thinking about passenger miles in a
19:10one-hour system
19:12so those are the two models so do you
19:14try to move a lot of people or you're
19:15trying to move
19:16a single person
19:19so let's think about that how long would
19:23it take you to
19:24move one person in the
19:27one person the quarter mile in the
19:29ferrari
19:30well that's just 10 seconds
19:33for a quarter mile how will take you how
19:35how how long would it take you to move
19:37one person in the bus a minute so the
19:39ferrari is much better than the bus in
19:41that model
19:43how long would it take you to move 54
19:46people
19:47a quarter mile well it takes a minute
19:50for the bus fill the bus
19:52up maybe there's some filling time and
19:53draining time fill the pipeline drain
19:55the pipeline
19:55but assume that that's a free thing
19:59and by the way assume you can go back
20:00instantly just for this way back
20:02instantly okay
20:03so in the case of moving one person or
20:06you certainly want the response time or
20:07latency of the ferrari
20:09but in terms of moving 54 people boy
20:12it's one minute to take a bus full of
20:13people a quarter mile
20:15how long will it take it would take
20:16sounds like i'm counting on
20:1827 10 seconds that's a lot worse than
20:21one minute
20:23so that's the throughput so the bus wins
20:25in terms of throughput
20:27the ferrari wins in terms of response
20:29time and these are important as you're
20:30thinking of
20:31response times of our system let's
20:33actually look at this
20:34and we talk about the array remember my
20:36server rack and array i've got this
20:37array now
20:38room full of computers i've got a local
20:41system
20:42i've got a rack and i've got my array
20:44and as i look at this there it's a
20:45really interesting little chart here
20:48i've got one rack in a rack and 30 racks
20:50in my array that's let's go to the high
20:52end here
20:53one server on my local computer 80
20:56max this number out 80 in my rack and
20:592400
21:00in my array eight cores
21:04640 cores 19 000 cores huh
21:07pretty juicy right don't you want this
21:08in your basement that'd be amazing
21:10look at this though look at this number
21:11so the numbers 16 gb for dram
21:141280 and 38 000.
21:18okay dram capacity
21:21disk capacity for tubby 320 and 9600 boy
21:25that's a huge data file that's very good
21:27okay nine pebby bytes right tabby and
21:30pepe bytes 9.6 pebby bytes
21:33now let's look at this following i have
21:34got dram latency
21:36so that's the time for one byte to be
21:38written to dram in terms of microseconds
21:41and disk latency one byte we written to
21:43disk look at the difference
21:46this by the way means this this means
21:48local rack array what this says is
21:50i'm gonna not only have a distributed
21:52file system but distributed memory
21:54system
21:55meaning i could write to another
21:57computer's dram
21:58if you allow for that in the of you have
22:00an os a distributed operating system as
22:01well
22:02that allows for that that's incredible
22:04it means that i can think of
22:05how much ram my computer can run not
22:07just what i locally have which is
22:0916 gb but i could think of my computer
22:11being able to have 38
22:13400 gibby pretty incredible so that's
22:16important you understand this idea that
22:18i'm going to be able to read and write
22:19not only to my local drive to my local
22:21memory but to my
22:23arrays any any drive on my array or any
22:27memory
22:28in my array as well that's important so
22:30now let's
22:31let's let's actually look at this and
22:32green is better and red is bad here
22:35so if i were to now look at the latency
22:38to write a single byte
22:39well remember disk is like going to
22:41andromeda
22:42so to write a single byte here is
22:4610 000 microseconds
22:49basically 10 milliseconds okay how about
22:53if i'm rather than writing to my own
22:55disc
22:56what if i could write to your when i say
22:58yours and other computers
23:00memory in my sam array your ram
23:03what if i could write it right to your
23:04ram if i had stuff that i can't
23:06you know here's my capacity of my here's
23:08my capacity
23:09of my of my disk four time bytes well
23:12let's look at this i've got dram
23:16of 38 tabi bytes
23:20so look i have got four versus 38.
23:23i if i had a thing that's oh you know
23:25it's like three teddy bytes
23:26well it's more than my ram has but
23:29certainly not more than my array has
23:31so really interesting conversations when
23:33you think of using ram
23:35as a disk my network ram as a disk
23:38really interesting
23:38and it is faster look at this by a lot
23:42to write to your ram rather than right
23:44to my disk as long as i can fit it
23:46all in that or maybe not just one person
23:49just one
23:49not just one machine but a collection of
23:51that really interesting if the os can do
23:53that for you
23:53and now let's look at how about if i
23:56have to just stream i can just stream
23:58data just
23:59stream stuff not just one but throughput
24:03higher bandwidth of another throughput
24:05here to local disk
24:07okay so this is now this by the way this
24:10is
24:11i want smaller numbers if i talk about
24:12latency i want a smaller number
24:14in bandwidth i want a higher number this
24:17is meg
24:17megabytes per second thinking about that
24:21so what am i doing here i am
24:25writing to my own disc
24:29twice as fast as i can write to your
24:32dram so remember thinking all right i
24:35can write to my own disk
24:36your disk my ram your ram certainly
24:38right into my ram is better but if i
24:39have a lot of data
24:41and i have to be just streaming it out
24:42there it makes more sense to write to my
24:45disk versus your
24:46ram and certainly the worst is your disk
24:49right that's even worse it takes all the
24:51penalty so it's interesting it's a
24:53really interesting conversation to have
24:54this
24:55last slide on this particular lecture to
24:57think that
24:58you might i mean the whole it's like
25:00you're like mind-blown
25:02where you can think of a system an os
25:04handling and allowing a particular
25:06machine
25:07to not just write to its own ram and its
25:08own disk which is what you thought of
25:10all along
25:11but to write to someone else's that's a
25:13distributed file system
25:14someone else's disk that's not so crazy
25:16your disk is on the network somewhere
25:17that's fine
25:18but also someone else's ram to think
25:20what's where what's the ram that i have
25:21access to
25:22it's all the ram in my array that's
25:24pretty interesting
25:25and that sometimes you want to write to
25:27my ram to to ram that is
25:29far away rather than your own ram or
25:31rather than uh not in your own random if
25:32you can that's your
25:33your old ram is going to be the fastest
25:34but rather than write to your own disk
25:36or someone else's ram right to their ram
25:38pretty incredible so i hope this opens
25:40your eyes a little bit about what
25:41they're doing in these systems to get
25:43the performance that they're getting
25:44now you remember spark was do it in ram
25:47look at all the ram you think about
25:48array having all the ram being able to
25:49be used
25:50and no one's ever touching disk at all
25:52except to initially probably read the
25:53file
25:54at the beginning of the day or being in
25:55the job and write the file at the end of
25:57the job but all the intermediate
25:58competition was done
25:58in ram that's why spark is better than
26:01mapreduce
26:02we'll see you at the next video

About this transcript

This page contains the full transcript of [CS61C FA20] Lecture 37.2 - Data Centers, Cloud Computing (WSC): Warehouse Scale Computers by CS 61C Departmental, generated from the public captions YouTube serves with the video. The transcript has 5,502 words across 881 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.