YT YouTube2Text

YouTube2Text → Transcripts → [CS61C FA20] Lecture 37.3 - Data Centers, Cloud Computing (WSC): Power Usage Effectiveness (PUE)

[CS61C FA20] Lecture 37.3 - Data Centers, Cloud Computing (WSC): Power Usage Effectiveness (PUE) — Transcript

by CS 61C Departmental · 3,928 words · 657 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00in this final video let's talk about
0:04power usage effectiveness thinking about
0:06optimizing for
0:07power so one of the things
0:10i'll tell you a story a long time ago i
0:12had someone from
0:14twitter visit our cs10 class it's about
0:18five or ten years ago
0:20and he talked about rafiki krikorian was
0:22his name
0:23and one of the things he did in his
0:24lecture is he talked about the
0:26difference in workload variation between
0:27when twitter's at the quietest
0:29or when it's at the most busy he said
0:31it's an interesting idea he said
0:34the difference was a massive scale the
0:35difference between the lowest time where
0:37twitch is not being used at all people
0:38mostly asleep around the world by the
0:39way the fascinating he said
0:40he said like you watch these usages
0:42follow with
0:43when people are awake around the world
0:45and as twitter had more and more
0:47folks in asia using it all of a sudden
0:49they saw that the load stopped going
0:50down it was
0:51mostly asleep because people in north
0:52america are asleep and then people in
0:54europe and asia where
0:55so you kind of get this world this
0:57little cycle thing of it
0:58he said when the world cup is happening
1:00unbelievable activity on twitter
1:02when um at the time madonna or justin
1:06bieber were tweeting and following each
1:08other
1:08he says like the worst thing for twitter
1:10is like
1:11it's a very funny story he said the
1:13worst thing we would ever have
1:15is if justin bieber and madonna were at
1:17the world cup
1:18and brazil's in it by the way that would
1:20be very also bad because there's a ton
1:21of people
1:22brazil brazilian users in twitter and
1:24brazil's in it and brazil
1:26hits the winning goal and then madonna
1:29and justin bieber
1:30are each tagging each other and
1:32commenting on that winning goal and
1:34selfies with the winning goalie or
1:36something
1:36so like we couldn't even imagine what
1:38how our system would explode because
1:39it's all about the who's following whom
1:41and it's just
1:41crazy what they've seen as they did this
1:44study on workload of variation that so
1:46their peaks and valleys were very
1:48different much farther than a factor of
1:49two
1:50but what they've noticed in some of the
1:52power usages
1:54of these of these warehouse scale
1:55computers is that there is enough
1:57distribution of
1:58just people are always hitting on amazon
1:59always sitting on these services like
2:01piazza and other things in facebook
2:02they're always hitting these services so
2:03that
2:04the amount of variation is only a factor
2:05of two even though twitter was saying
2:07for their particular servers it was much
2:09more attacked for 10 or 20 or more a
2:10hundred even in the worst case
2:12they're saying only really a factor of
2:14two differences in workload variation
2:16so that's a big difference that once you
2:17have a lot of more maybe
2:19a lot of large numbers remember twitter
2:21is just one service
2:22but really these servers are lots of
2:23different services going on at once
2:25so maybe that kind of is smoothing the
2:27numbers down to have that
2:28not at a peak at peak time so handling
2:31workload variation is an important thing
2:34so what are the in what's the impact of
2:35all the following things latency
2:37bandwidth failures varying workloads on
2:39works
2:40warehouse scale computer software
2:43well first of all you've got to know
2:45where to place data
2:47in your array um to get good performance
2:49that we talked about that before
2:50um you don't put at the edges and if
2:54if you have one company or one service
2:55using it put it close together close
2:57together near a switch
2:58don't put it well i'll put you on a
2:59server over there through 19 different
3:01switches to get to one end half the data
3:03is there and have no put it as close as
3:04you can tighten it again
3:05same machine if you can all those things
3:07um you've certainly got to handle
3:09failures gracefully we talked about how
3:10mapreduce handles it if machine goes
3:12down
3:13we just send it the job to somebody else
3:14that's no problem but you have to
3:16generally deal with know this you also
3:17know how to flag them whatever she goes
3:19down
3:19you as someone who's working at the i.t
3:21desk and that you need to know what
3:22drive is down failure you need to know
3:24when a drive fails but before the rate
3:26the raid array fails so
3:27that's that's something to catch very
3:28quickly to either power that down so
3:30that
3:31you so that or maybe if you get to it
3:33quickly enough
3:34often things happen in bursts because
3:36you replace all the computers at the
3:37same time you put all the hard drives at
3:38the same time
3:39and they often have failure modes all
3:40around the same time so when one fails
3:42you often say you know what
3:43let me replace all of them because i see
3:45them starting to fail
3:46so rather than just wait till they do
3:48fail you can often preemptively replace
3:49them because you're saying you know
3:50they're at about a lifespan
3:51they're all manufactured identically
3:53right shouldn't they all fail about the
3:54same time so that's what starts to
3:55happen
3:55one fails two fails all of a sudden
3:57there's a ton of them so they have a
3:58whole
3:59you know scale replacement and when they
4:00do that then you have the problem now
4:02they'll all need to replace again at the
4:03same time roughly
4:05you certainly must scale up and down um
4:07you want to power things down if you can
4:09uh or maybe just go idle and you see how
4:11bad it is so you think about
4:13power and what you should do in terms of
4:15your own costs to deal with scaling up
4:17and down as usage uh
4:18increases or decreases um there's
4:22just complex complexities of hierarchies
4:24of memory
4:25failure tolerance workload
4:27accommodations all that all that makes
4:29that
4:30the software development for warehouse
4:31scale computing really challenging
4:33much more challenging than a single
4:34computer where you as i mentioned before
4:36cores don't die
4:37i maybe they do but rarely really really
4:40rarely
4:41i mean your hard drives die but you know
4:43you'll have one drive die or that but
4:44you don't have a core dies
4:46you don't have a worker b die that
4:47happens all the time though you know
4:49where the network goes out or a computer
4:50dies
4:51those will happen all the time
4:52eventually machines die although they're
4:54certainly
4:54you know with solid-state drives you
4:56certainly have failures less often than
4:58that used to of their
4:59early days so here's a graph power
5:02versus
5:03server utilization and this is a really
5:05interesting graph but also a
5:07i wouldn't say scary but it's a it's a
5:08depressing graph
5:11this shows as the computer load grows
5:13from idle
5:14to 100 what's the power okay so it turns
5:17out that these systems
5:18are not just constant power that would
5:20be terrible to have a computer uh be you
5:22know
5:22idle use exactly the same power as at
5:24maximum time
5:26what they found in face look at this
5:28curve
5:29here's you know here's here's the factor
5:31that contributes a cpu dram disc
5:33interesting and other what contributes
5:34to the power usage of a system
5:37look at this uses almost half power
5:40when idle almost half power when idle
5:45and two-thirds power when 10 utilized
5:48look at this here's 10 utilization
5:51here's look at this
5:53two thirds when ten percent utilized
5:57ninety percent 50 here's fifty percent
6:00ninety percent power at fifty percent
6:03utilization
6:04um most servers by the way you want to
6:08keep in the sweet spot you don't want to
6:09have most servers be
6:10pinned because then what happens there's
6:11no room to breathe you want to be able
6:13to have
6:14most servers live in kind of the safe
6:16spot here
6:17and by the way the goal should be energy
6:19energy proportionality the peak load
6:21should be equal to the percentage energy
6:23so this should be a straight line okay
6:26that's not a straight line but it should
6:27be
6:27close to that you're nowhere near that
6:29so that's the goal here and
6:31people are trying to build systems more
6:32and more to get to that but it's hard
6:33it's very hard
6:36so power usage effectiveness this is a
6:39really interesting model and it's a very
6:41nice metric the overall efficiency is
6:45the amount of computational work
6:46performed
6:47divided by the total energy used in the
6:49process
6:51that's it total building power so pue
6:54is total total building power divided by
6:57it equipment power
7:00so power efficiency for warehouse scale
7:02computers nothing nothing
7:04not including the efficiency of servers
7:05and networking gear so
7:07a one means total building power divided
7:09by i t equipment power
7:10is a one when you're doing exactly right
7:13so
7:13here are numbers here's the best
7:16possible value the best possible value
7:18so
7:19essentially there's overhead right if
7:20you're thinking about this the higher it
7:21is it means you just have more waste
7:23there's just more not as much overhead
7:24if you're exactly perfectly lean you're
7:26computing exactly all the cost is for
7:28the it stuff not the overhead
7:32so look at this so this is very
7:34fascinating lawrence berkeley national
7:35laboratory did a survey
7:36of the pue of 24 different data centers
7:39in 2007
7:41and what they found was a remarkable
7:43variation so
7:44you know 27 24 data centers who knows
7:47how many different designs if you look
7:48at these
7:49are the different flavors well these
7:50kind of the same those are kind of the
7:52same that's maybe
7:54these are all the same design maybe this
7:55may be the same design who knows who
7:57knows whether people do same design
7:58who knows what other systems one might
8:00be in a hot area or versus a cold area
8:02who knows where they are how whether
8:03they have any other differences
8:05are these 24 total different designs or
8:07basically the same design with different
8:08variations of it
8:09and they found the average about 1.83
8:12the goal is to get to one
8:14it's very hard to get to one but there's
8:15some that are doing really well
8:17there's something about 1.2 1.3 1.4
8:20that's pretty incredible
8:21and there are some that are just crazy
8:22and efficient at three here so really
8:24interesting to think about
8:25power use usage uh pue i'm sorry pue in
8:29the wild
8:31where does it go where does all that
8:33excess power go when you're not spending
8:35time
8:35on on the actual compute remember the
8:38goal is to get it to be all the powers
8:40being used for the things for the it
8:42system things that are
8:43generating paying and paying the bills
8:45at the end of the day
8:46um so here's the service of networking
8:50which is the it equipment that's there
8:51now where else is it going
8:53well power distribution unit is a little
8:56bit of time
8:56there's transformers and switches a
8:58little lost there there's the ups we
9:00talked about before this is probably the
9:01ups that isn't
9:02local to every computer but that might
9:04be local to every computer as well
9:07look at this chiller
9:10cools warm water from the air
9:11conditioner and then there's a computer
9:13room air conditioner
9:15okay so you've got a computer room air
9:17conditioner
9:18moving air in and out and then that hot
9:20air has to be cooled
9:22to then send cool air back in and
9:24there's a chiller doing that for making
9:25sure if you've ever seen that condition
9:26there's two parts of that
9:27look at this look at the size of it it's
9:29like the half of the
9:31chart is just cooling
9:34incredible so what did google learn so
9:38google did this
9:39and wrote a wrote a report they said
9:41this is what we've learned
9:43how to be really smart about pue one
9:46careful airflow by the way whoever
9:48thought that if dan you're gonna be
9:49teaching computer
9:50architecture okay great you're talking
9:52about airflow
9:53what yeah sometimes you need to talk
9:55about airflow and cooling
9:57and temperature uh if that's the reason
9:58that both you're paying a ton of money
10:00and those are the reasons that that you
10:02you can't run something at a certain i
10:03mean
10:04i mentioned before the reason we can't
10:06have six gigahertz 28 core machines
10:09the current intel machine is a 28 core
10:112.5 gigahertz machine these are
10:13six gigahertz because i can't get that
10:14cool there's just no way to get heat off
10:16of that chip
10:17so it's a reality in terms of the
10:19constraints so we talked they talked
10:21about google talked about
10:21careful airflow handling don't mix the
10:24server hot air exhaust
10:26here's the server and the hot as cool
10:27air coming in and it takes the heat off
10:29of the server and then
10:30hot air is coming out don't mix it with
10:32cold air separate
10:33a warm aisle from a cool aisle which is
10:35really interesting it's almost like
10:38it's almost like uh like the stockade
10:40where you have a cow and the cow gets
10:41fed on this side
10:42and then the back side you just clean up
10:43on the waste end of it it's like this
10:45you kind of lock in the server
10:46two parts of it and there's no flow of
10:49air from the front air cold air coming
10:51in
10:51coming in pulling the heat off of the
10:53system and then the hot air i think of
10:55the exhaust as the garbage
10:56goes out the back and so keep those
10:58distinct
11:00also they learned very clearly that just
11:02having one big room
11:04i mentioned the olden days you might
11:05have been surprised if you saw
11:07inside the inside of a wsc you see these
11:10freight containers
11:11you might say why did they do that what
11:12they just have a big room and all these
11:14servers
11:14yeah cause it's really loud that's not a
11:16big deal but also this hot and cold air
11:18just mix
11:19and you can't get the the whole thing
11:21gets really really hot
11:22so they said if you localize it to this
11:25freight container then you can have a
11:26lot more channeling of air
11:28rather than air just swirling around and
11:29eddies and flows it's a lot easier to
11:31control the airflow in a smaller
11:33container so that's what they do that's
11:34the reason why they put them in there's
11:35a smaller
11:35you also can put them in and replace
11:37them and take a whole thing oh take it
11:38home
11:38here's a new one there's an old computer
11:40okay here's a whole new one we're
11:41replacing all the computers by just
11:42putting a new freight container in droop
11:43new freight container all new computers
11:45that's kind of a nice thing so
11:46somebody's obviously switching them but
11:47you just swap it in very quickly and all
11:49of a sudden you swap it very fast that's
11:50another advantage of that
11:52short path to cooling so little energy
11:54spent moving cold or hot air long
11:57distances
11:57so that means it's expensive and hard
12:00and inefficient to take hot air
12:02and move it all the way over here to
12:03cool it to send it all the way back
12:05so have a short path of cooling as much
12:06as you can localize cooling it's almost
12:08like data right you want to have the
12:09data
12:09near the compute center don't put one
12:11data over here one day over there and
12:12here's the cpu no no have the data close
12:14to it
12:14same idea here if i short path to
12:16cooling and i mentioned before that's
12:17the third build i kind of
12:18i skip orders but keeping the servers
12:20into the containers helps the airflow
12:22you can you get to control the airflow
12:23because the airflow isn't just random
12:24all around the room also they learned
12:28which is a funny thing i didn't i
12:29wouldn't have realized this
12:30that you can elevate the cold aisle
12:33air temperature normally the cold aisle
12:35means okay how do i get that out of here
12:36but then like
12:37how how cold is it normally walk into a
12:39server room it's pretty warm
12:41they've learned that you can actually
12:42keep those server rooms warmer
12:44and not have machine failures the worry
12:46is if you have a higher elevated
12:48temperature around the machines
12:50you're gonna have failures you're going
12:51to have the computer oh something things
12:53will overheat if you ever have
12:54folks who live in very very hot areas in
12:56the caribbean their computers fail more
12:57often because it's just a hotter
12:58temperature
12:59just it's harder to cool those things so
13:01they found that it's actually okay to
13:03keep the hot the
13:04cold aisle temperatures in the 80s
13:06rather than the 60s i used to walk into
13:08soda hall soda hall and corey hall into
13:10their cool into their
13:11server rooms there's a couple server
13:12rooms that used to live on the fifth
13:13floor and i'd walk in
13:14and be freezing in fact they actually
13:16moved a couple of
13:18they moved the lab inside the inside
13:19there and i would i could be freezing
13:21working on the computer
13:22with the air conditioners in the other
13:23room but it was just so cold in near
13:25that system
13:26because they thought well you got to
13:27keep them cold they learned you don't
13:28have to keep them that cold
13:29you can actually keep them a little bit
13:30warmer i said
13:32the reliability is okay if the service
13:33would run hotter and i mentioned this
13:36earlier use free cooling
13:38cool warm water out by evaporation
13:40cooling towers so
13:41right right there and also put it in a
13:43moderate climate i've mentioned that
13:44before
13:45and maybe even draw water in it to help
13:46you the cooling system if you can
13:48and i mentioned this before a per server
13:5012 volt
13:51ups rather than a a warehouse scale
13:54ups in the basement place a single
13:56battery per server board
13:58increases the efficiency from 90 to 99
14:00so huge wins on that
14:01having a little battery for everybody
14:03and
14:04don't just go it alone measure and
14:07estimate pue publish it
14:09improve operations share the best
14:11practices you know you know you want to
14:12here here's google here's facebook
14:14here's amazon don't talk no
14:16talk to each other and the best
14:17practices talk to each other let's have
14:18conferences around
14:20wsc and let's figure out what the best
14:22ideas are to help all of us it's really
14:23nice idea rather than
14:24all go it alone and never share any
14:26value value information i mean certainly
14:27there's some
14:28corporate secrets about these things but
14:30like just it's kind of a nice idea that
14:31open the open-minded spirit to share
14:34the best practices across across
14:35different companies nice nice model
14:38this is an interesting computer the news
14:40piece normally computer news happens
14:41during the live sessions or doing
14:42you know welcome to cs6216 computer the
14:45news
14:46i want to put this the end because now
14:47we're talking about power
14:49in 2011 google disclosed it started
14:52people
14:52somebody revealed how much power google
14:55was actually using somebody says you
14:56know tell me how much power they went
14:58out they went out and did some
14:59investigation to find out how much power
15:00google was using in all of its service
15:01and all of its servers
15:03and it said it continuously uses enough
15:06power to
15:07enough enough electricity to power 200
15:10000 homes
15:11which blew people's minds like that's a
15:13small city
15:14you're telling me all your servers could
15:16power 200 000 homes
15:18but here's what it says it says by doing
15:20so it makes the planet greener
15:21why well you might argue that's an
15:24interesting case
15:26before without google without the
15:27ability for google to do it not the ad
15:28part but google to be able to give
15:30search i might have to call people the
15:31phone i might actually go there i might
15:32have to go places
15:33are you open if i can't find out i have
15:35to go places so think about all the
15:36possible time that you're saving by
15:38using google rather than driving to your
15:40library or driving to buy something you
15:41know
15:42all those services google provides they
15:44say well you know what
15:45search cost per day the same as running
15:47a 60 watt light bulb for three hours
15:50so turn your six foot light bulb on for
15:52three hours and turn it off
15:53and then that's the same as searching
15:55per day per person
15:56it's a model they still got flack for
15:59that
16:00so google then said this is just the
16:02google side
16:04and this is some news in 2018 over the
16:06course of 2017 across the globe for
16:08every kilowatt hour of electricity we
16:09consumed
16:10we purchased kind of like carbon trading
16:13in a way
16:14a kilowatt hour of renewable energy from
16:16a wind or solar farm that was built
16:17specifically for google
16:19that makes us the perf public cloud and
16:20company of our size to have achieved
16:22this feat
16:22so kind of trading that in that sense so
16:24that's an interesting model
16:27this is a curve this is a graph as part
16:30of the community in the news
16:31that says cumulative corporate renewable
16:33energy purchased in the united states in
16:34europe and mexico
16:36march 2018 and you see that google is
16:38far outpacing anybody else in terms of
16:40the big
16:40people running services but i haven't i
16:41didn't microsoft also runs warehouse
16:44scale computing as well
16:45as well as apple but it's kind of
16:46interesting you can see who's running
16:47these ibwc look walmart here dow
16:50chemical
16:51interesting right not just the big kind
16:53of three or four or five
16:54so google's certainly buying a lot more
16:56um
16:57renewable energy so that's a great kind
16:59of offsets like offset right to offset
17:01all the work that they're doing for
17:02there
17:03we're done oh my gosh in summary
17:06parallels is one of the great ideas in
17:08computer architecture number four in
17:10fact
17:10it applies to many levels in the system
17:12from instructions
17:14from apparel and gate down to the up all
17:17the way up to a warehouse scale
17:18computing pretty incredible that you
17:19learned all about this in one class
17:21remember post pc era local drive
17:24local local watch computing device has
17:28the front end interface for it
17:29really wonderful voice interface amazon
17:31alexa all that
17:32but the back end the cloud does all the
17:34work that back end the warehouse scale
17:36computer has to deal with failures
17:38varying workload
17:40varying hardware latency
17:43certainly sensitive to cost energy
17:45efficiency you put them in the right
17:46place in the in the
17:47in the country you try to draw them you
17:49try to get cheap land you try to deal
17:50with the
17:50cooling can you deal with even cooling
17:53without ever
17:54i'm going to show you two videos after
17:55this i'm linking to two videos as part
17:57of this
17:57i want you to watch i believe it's a
17:58google and a an amazon
18:00i believe it's a google it's like google
18:02on a facebook uh wsc i want you to watch
18:04these videos which are just little like
18:06news videos about what goes on inside
18:07these i want you to make sure you watch
18:08that as well
18:09um and warehouse scale computers support
18:12many applications that we have come to
18:13depend
18:13on so this is how it works this is our
18:16little taster 161c i hope you enjoyed it
18:18i'll see you next time

About this transcript

This page contains the full transcript of [CS61C FA20] Lecture 37.3 - Data Centers, Cloud Computing (WSC): Power Usage Effectiveness (PUE) by CS 61C Departmental, generated from the public captions YouTube serves with the video. The transcript has 3,928 words across 657 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.