Inside Azure datacenter architecture with Mark Russinovich – BRK3347


WAS LAME.GOOD AFTERNOON,EVERYBODY.>>Audience: GOOD AFTERNOON!>>THERE WE GO.HOW IS IGNITE GOING?HOW MANY FIRSTTIMERS?OKAY,QUITE A FEW OF YOU.WELCOME TO IGNITE.MY NAME IS MARK RUSSINOVICH,CHIEF TECHNOLOGY OFFICER OF
MICROSOFT AZURE.HOW MANY PEOPLE OF HEARD OFAZURE,BY THE WAY? WHEN I FIRST STARTED DOING THIS TALK,THAT WAS A REAL QUESTION.WE’VE COME A LONG WAY.I’M GOING TO SHOW YOU SOME OF
THE LATEST INNOVATIONS WE HAVE,GIVE YOU A LOOK AT THINGS THAT
ARE COMING,AND ALONG THE WAY, FRAME THAT INOVERVIEW OF THE
AZURE ARCHITECTURE.SO THIS WILL TAKE YOU INSIDE
INTO VARIOUS ASPECTS OF THE AZURE PLATFORM.HERE’S AN AGENDA. IT’S A TON OF MATERIAL.I THINK ONE OF THE MOST
DIFFICULT THINGS FOR ME ABOUT
THIS TALK IS THAT I START OUT
WITH A WHOLE COLLECTION OF DEMOS ANDCOOL FEATURES,INTERNALINFORMATION, THAT WOULD END UPBEING LIKE THREE OF THESE IN
ORDER.SO I’VE GOT TO TRIM THEM ALL
DOWN TO JUST A FEW. YOU CAN SEE IT’S PAINFUL CUTTING IT DOWN.I’M GOING TO BE MOVING VERY QUICKLY.I’VE GOT TEN DEMOS AND THERE’S SOME COOL SHOCK AND AWE DEMOS.BUT I’M GOING TO GIVE YOU
BACKGROUND ON OUR DATACENTERS,TALK ABOUT OUR DATACENTER
ARCHITECTURE,OUR PHILOSOPHY ABOUT HOW WE PLACETHEM,WHERE
WE’RE GOING,FUTURE-LOOKING
ASPECTS,GET INTO PHYSICAL
NETWORKING,HOW WE CONNECT THE
DATACENTERS TOGETHER AND WHAT WE
DO WITHIN THE DATACENTER WITH
NETWORKING,TALK ABOUT LOGICAL NETWORKING,THE VIRTUAL
NETWORKING CAPABILITIES,GO INTO
THE SERVER DESIGN,A HISTORY OF
SERVERSFROM 2010 TO NOW,GO INTO OUR COMPUTEPLATFORM, WHICH INCLUDES HOWWE DEPLOY
VIRTUAL MACHINESES AS WELL ASCONTAINERS. I’LL SHOW YOU SOME REALLY COOL
INNOVATIONS, INCLUDING SOME THAT
WE ANNOUNCED THIS WEEK.TALKING ABOUT AZURE STORAGE, ONE
OF THE FUNDAMENTAL BUILDING BLOCKS OF ANY CLOUD APPLICATION. AND WE’VE DONE A NUMBER OF
THINGS THIS YEAR TO IMPROVE THE
SCALABILITY OF OUR STORAGEPLATFORM,AS WELL AS IMPROVE THE
PERFORMANCEOF OUR DISK PLATFORM.SO I’LL COVER THAT.ANDFINALLY, A BRIEF LOOK AT OUR
DATA PLATFORM. I’M GOING TO SHOW YOU SOME
THINGS THAT ARE COMING THAT
SHOULD EXCITE YOU.AND THEN TALK A LITTLE BIT ABOUT COSMOSDB.SO LET’S GO AHEAD AND GET STARTED BY TAKING A LOOK INSIDE OF OUR DATACENTERS.AND LIKE I SAID, I’M GOING TO
START WITH HOW WE PLACE DATACENTERS.WE DIVIDE THEWORLD INTO GEOGRAPHIES.THESE USED TO BE LARGE AREAS
LIKEWESTERN EUROPE,NORTH
AMERICA. BUT MORE AND MORE IT’S BECOMING A COUNTRY.AND YOU’LL SEE OUR GEOS MAPPING
TO COUNTRIES NOW.WHAT WE BELIEVE,OUR PHILOSOPHY
ABOUT WHERE WE DECIDE TO CREATE
ONE OF THESE GEOS AND HOW WE
DEFINED BORDERS IS CUSTOMERS ARE
GENERALLY COMFORTABLE HAVING
THEIR DATAANYWHERE WITHIN THAT
BORDER OF THATGEO.SO WITHIN A GEO,WHERE WE PLACE
DATA CENTERS, WE CALL THOSE
REGIONS.THE REGION IS DEFINED BY A
LATENCY ENVELOPE OF TWO
MILLISECONDS.WE PLACE THEM CLOSE ENOUGH
TOGETHER THAT DATA CAN MOVE WITH
ABILITY TO SYNCHRONOUSLY REPLICATE DATA WITHIN A REGION.THEN WE ALSO GO INTO REGION PAIR
ARCHITECTURE WITH ALMOST ALL OF
OUR REGIONS.THERE’S ONLY ONE EXCEPTION RIGHT
NOW AND THAT’S SOUTHAMERICA.THE BRAZILIAN REGION,IS PAIRED
WITH SOUTH CENTRAL U.S.EVERYWHERE ELSE IN THE WORLD THERE’S TWO REGIONS.WE WANT TO CREATE THE ABILITY
FOR CUSTOMERS TO DR FROM ONE
REGION TO THE OTHER.WE SEPARATE THEM BY HUNDREDS OF
MILES SO THEY CAN TOLERATE LARGE
EVENTS LIKE THE POWER GRID
FAILURE,A HURRICANE, WHICH IS
TOP OF MIND FOR A LOT OF PEOPLE
THESE DAYS.AND WE ALSO,TO ENSURE HIGH
AVAILABILITY OF THOSE REGION
PAIRS,WE UPDATE SOFTWARE TO
THEM INDEPENDENTLY.WE DO ONE AND THEN THE OTHER IN
CASE THERE’S A PROBLEM WITH THE
UPDATE BRINGING DOWN PART OF A
REGION.WE DON’T BRING DOWN BOTH REGION
PAIRS AND CUSTOMERS CAN FAIL
AWAY.ONE OF THE ASKS THAT WE’VE
GOTTEN FOR A LONG TIMEIS
RESILIENCY TO LOCALIZED FAULTS
BEEN — WITHIN A REGION,LIKEA
POWER OUTAGE.WE’VE HAD SOME OF THOSE IN OUR
PAST.AND SO TO HAVE CUSTOMERS BE ABLE
TO PROTECT AGAINST THOSE,WE
STARTED TO ROLL OUT SOMETHING
CALLED AVAILABILITY ZONES.AVAILABILITY ZONES ARE TAKING A
REGION AND SPLITTING IT INTOSEPARATE — THREE SEPARATEISOLATED DATACENTER AREAS WHERE
THOSE DATA CENTERAREAS ARE —
HAVENON-CORRELATED FAULT MODES.BY THAT I MEAN THEY’REON
INDEPENDENT ELECTRICITY,POWER,WATER,AND THEY’RE INDEPENDENT
PHYSICAL FACILITIES. SO IF THERE’S A FLOOD IN THE
DATA CENTER IT’S ONLY GOING TO
AFFECT ONE,NOT THE OTHER TWO.THE REASON WE GO IN WITH THREE
IS WE WANT CUSTOMERS AND US TO
BE ABLE TO CREATE PERSISTENT STORAGESERVICES THAT REPLICATEDATA. AND TO REPLICATE DATA WITH HIGHDURABILITY, YOU NEED A VOTE ON
WHAT IS THE MOST DURABLE DATA.THAT MEANS YOU NEED AT LEAST
THREE TO ACHIEVE THAT WHERE YOU
CAN LOSE ONE AND STILL BE ABLE TO WRITE DATA WITH THE VOTE OF
TWO OF THREE,WITH THE THIRD ONE
NOTPARTICIPATING. CUSTOMERS WILLSEE THREE ZONES.REGIONS COULD HAVE POTENTIALLY
MORE THAN THAT.WE’RE DOING A LOGICAL TO
PHYSICAL MAPPING.HERE’S A HYPOTHETICAL REGION THAT HAS FOURZONES.AND TWO CUSTOMERS WITH DIFFERENT
SUBSCRIPTIONS MIGHT HAVE
DIFFERENT MAPPINGS ON THEIR ZONES THAT THEY SEE TO THEPHYSICAL ZONES UNDERNEATH.NOW,THAT’S KIND OF THE LOGICAL
VIEW OF HOW WE PLACE DATA
CENTERS.HERE’S THE PHYSICAL VIEW.I’VE GOT A FEW PICTURES OF SOMEOF THE DATACENTERS IN OUR
REGIONS. HERE’S TWO DATACENTERS.THERE ARE TWO DATACENTERS THAT
CONSISTOF FOUR INDEPENDENTCO-LOS,THIS IS IN QUINCY, WASHINGTON,NEAR THE COLUMBIA
RIVER. AND THATREFLECTS OUR GOAL TO
POWER OUR DATA CENTERS OFF
RENEWABLE ENERGY SOURCES,LIKE
THE COLUMBIA RIVER. HERE’S ANOTHER LOOK AT CHEYENNE, WASHINGTON.SO YOU CAN SEE A SIMILAR DATA
CENTER ARCHITECTURE THERE.THIS ARCHITECTURE ISCALLED
DC2015.DC2015. IT’S A FEW YEARSOLD.THIS IS ONE OF OUR LARGERREGIONS,OUR REGIONS THAT WE ARE
CREATING TO BASICALLY BE ABLETOEXPAND INDEFINITELY IN TERMS OF
THE FOOTPRINT THAT WE CAN PUT
THERE.LOTS OF LAND,LOTS OF POWER. HERE’S A LOOK INSIDE THAT
DATACENTER. HERE’S A HALLWAY.IT KIND OF LOOKS LIKE THE DEATHSTARHALLWAY,BUT IT’S RELATIVELY BORING.WHAT’S WEIRD IS WALKING INSIDE
ONE OF THOSE AND HAVING
CONTRIBUTED TO THE SOFTWARE THAT’S RUNNING ON THESE.I HAVE NO IDEA WHAT’S RUNNING ON
THOSE SERVERS. IT’S DOING WHO KNOWS WHAT.HERE’S A LOOK AT DUBLIN,IRELAND,ANOTHER ONE OF OUR
LARGER REGIONS.YOU CAN SEE THERE’S TWO
DATACENTERS IN THIS PICTURE.AND THERE’S SLIGHTLY DIFFERENT ARCHITECTURE.WE’VE BEEN BUILDING OUT IN
DUBLIN FOR THE LAST EIGHT YEARS.DATACENTER ARCHITECTURE HAS CHANGED SEVERAL TIMES SINCE
THEN. HERE’S ANOTHER LOOK AT DUBLIN.YOU CAN SEE ANOTHER DATACENTER
FACILITY IN ADDITION TO THETWO
WE WERE JUST LOOKING AT.THIS IS FROM A DIFFERENT ANGLE.THE LONG ONE ACROSS THE TOP. WE ALWAYS ITERATE ON DATACENTER DESIGN.THE REASON WE DO SO IS TO OPTIMIZE EFFICIENCY OF SERVERS,OR TRY TOGET WASTEDELECTRICITY
DOWN. YOU WANT TO MAXIMIZE THAT
ELECTRICITY SO THE SERVERS ARE
CHEAPER TO RUN. ONE OF THE WAYS WE’RE LOOKINGAT WITH MICROSOFT RESEARCH ON
RUNNING DATACENTERS EXTREMELY
CHEAPLY IS BY PUTTING THEM ON
THE OCEAN FLOOR. THIS ISA PROJECT PEOPLE HAVE
HEARD OF.SO QUITE A FEW OF YOU,A COUPLE YEARS AGO THERE WAS VERSION ONE, WHICH MADE A STORY IN THE NEW
YORK TIMES.MICROSOFT RESEARCH TOOK A VERY
SMALL CYLINDER,PUT ONE RACK IN
IT AND DROPPED IT IN 30 FEET OF
WATER FOR A MONTH,THENPULLED IT UP AND SAID STILL WORKING,COOL. WE DIDN’T LEARN A TREMENDOUS
AMOUNT FROM THAT.BUT WHAT WEHOPE TO LEARN ARE
SOME ASPECTS OF HOW WE CANOPTIMIZE THE ENERGY EFFICIENCY
AND MAXIMIZE THE LIFETIME OF THE
SERVERS.I’LL GET TO THAT IN A SECOND.WHAT YOU WERE JUST SEEING THEREIS THE12–METER-LONG TUBE WITH
12 RACK OF SERVERS IN IT.AND THAT IS THEN GOING TO BE DROPPED TO THE OCEAN FLOOR.YOU CAN SEE THIS GANTRY,WHICH
IS GOING TO BE FLOATEDOUT ONE
KILOMETER AND DROP THE CYLINDER
TO THE OCEAN FLOOR. IN ABOUT 100 FEETOF WATER.THIS IS OFF THE COASTOFSCOTLAND,NORTHERN SCOTLAND.ONCE IT’S DROPPED,WE NEED TOCONNECT IT TO ELECTRICITY AND NETWORKING.YOU CAN SEE THIS IS THE TUBE
THAT COMES FROM LAND THAT WE CONNECT UP TO THECYLINDER.THESE CYLINDERS CAN GO FROM
BEING MANUFACTURED SAYING WE
WANTONE TO BEING DROPPED IN 30
DAYS.THAT’S WHAT WE DID IN THIS CASE.THE CYLINDER WAS MADE IN FRANCE.WE HAD ITSHIPPED TO SCOTLAND AND DROPPED ON THE FLOOR 30 DAYS LATER.THE REASONS WHY WE THINK THIS CAN BE INTERESTINGIS THAT THIS
IS A LIGHTS-OUT OPERATION.BESIDES THE OBVIOUS THING THAT
IT’S COOLED BY THE OCEAN WATER,AND IT’SLIQUID-COOLED, SO THOSE
SERVERS WERE ENCASED.WE HAVE LIQUID WATER FROM THE
OCEAN COMING IN,COOLING THE
SERVERS.IS THAT WE’RE NOT TOUCHING THE
SERVERS. IT’S A LIGHT-OUT OPERATION.THE THEORY IS IF HUMANS AREN’T
MANIPULATING THEM,IF THE
TEMPERATURE REMAINSCONSTANT, IF
THE WATER VAPOR IS EXTRACTED FROM THE AIR AND THE AIR IS
EXTRACTED FROM THE CYLINDER,SO
IT’S BASICALLY A VACUUM,AND
IT’S ISOLATED FROM THE OUTSIDE
WORLD,THOSE SERVERS WILL LASTLONGER.THIS EXPERIMENT IS GOING TO LAST
THREE YEARS.THEY’RE SITTING ON THE OCEAN
FLOOR.WE EXPECT TO HAVE A PRETTY GOOD
VIEW OF SERVER LIABILITY AFTER A
YEAR OF OPERATING THIS.NOW,WE ARE CONCERNED ABOUT POTENTIAL IMPACT TO THE
ENVIRONMENT. WE’VE DONE ANALYSIS AND IT SHOWS THAT WE COULD PUT LOTS OF THESE ON THE OCEAN FLOOR AND WE’RE NOT GOING TO AFFECT THE LOCAL
SURROUNDINGS WITH THE HEAT
THAT’S DISSIPATED FROM THESE.THERE’S JUST SO MUCH ABILITY FOR
THE OCEAN TO ABSORB SO MUCH
HEAT.BUT NEVERTHELESS,WE DO WANT TO
UNDERSTAND IF THERE’S GOING TO
BE ANYIMPACT.SO WE’VE MOUNTED CAMERAS DOWN
THERE.AND THIS IS ACTUALLY A FEED FROM
ONE OF THECAMERAS.SO THIS IS PUMPINGIMAGES UP
INTO AZURE MACHINE LEARNINGPIPELINE,WHICH IS USING A FISHCLASSIFIER THAT A MARINE BIOLOGIST IS LOOKINGAT.AND WE’RE TRYING TO TRACK THE
HEALTH OF THE ECOSYSTEM AROUND
THESE TANKS.THERE’S A LIVE FEED ON THE
WEBSITE FORNADIC, BUT IF YOU GO
THERE,THE CAMERA IS COVERED BY
SEA WE’D.SEAWEED.WE HAVE TO HAVE PEOPLE GO DOWN
THERE TO CLEAN THE SEAWEED OF
THECAMERASOCCASIONALLY. [ LAUGHING ]
>>Mark: LET’S TAKE A LOOK AT
PHYSICALNETWORKING.WHEN IT COMES TO PHYSICAL
NETWORK,THE ARCHITECTURE IS
DESIGNED TO CONNECT OURREGIONS
TOGETHER OVER OURWAN. YOUCAN
SEE THE FACT THAT WE CAN CONNECT
TO CONSUMERS,ENTERPRISED THROUGH EXPRESS ROUTE DIRECTCONNECT, INTERNET EXCHANGES.AND THEN WE’VE GOTA REGIONAL
NETWORK WHICH CONNECTS THE DATA
CENTERS AND THE REGION TOGETHER
AND THEN TO THE WAN.THIS NETWORKING INFRASTRUCTURE
SUPPORTS ALL OF OUR NETWORKING
SERVICES.AND YOU CAN SEE A BUNCH OF
NETWORKING SERVICES THAT ARE
BOTHLOGICAL AND PHYSICAL THEREACROSS THE BOTTOM. AND THAT SET OF SERVICES
CONTINUES TO EXPAND.WE’VE GOT A COUPLE WE ANNOUNCED
HERE ATIGNITE.THAT DARK FIBER BACKBONE IS ONE
OF THE LARGEST IN THE WORLD.WE THINK IT’S ONE OF THE TWO
LARGEST DARK FIBER NET WORK BACKBONES CONNECTED DATA CENTERS TOGETHERIN THE PLANET.YOU CAN SEE A MAP,TENS OF
THOUSANDS OF MILES OF CABLE
CONNECTING OUR DATACENTERS TOGETHER,INCLUDING SOME VERY
EXPENSIVE PROJECTS THAT WE HAVE
GONE ON WITH OTHER COMPANIES TO
LAY CABLES ACROSS THE PACIFIC
AND ACROSS THE ATLANTIC JUST
VERY RECENTLY,ABOUT A YEAR AGO
WE ANNOUNCED THEMAREA CABLECONNECTING SPAIN,SOMEWHERE IN
SPAIN,THE COAST OF SPAIN,AND
VIRGINIA.MOST OF THE CABLES IN THE
ATLANTIC CONNECT UK AND NEW
YORK.SO THIS PROVIDES AN EXTRA LEVEL
OFREDUNDANCY.PLUSIT’S 160 TEARBITS OF
CAPACITY,MAKING IT THELARGEST
CABLE CROSSING THE ATLANTIC.WE HAVE 3500 PEERING POINTS WITHPARTNERS,103 POINT OF PRESENCE OUT THERE.AND THOSE POINTS OFPRESENCE, I’LL TALK LATER ABOUT THAT.ONE OF THE WAYS THAT WE CAN MAKE
A REGION SCALABLE ISBY NOT
DIRECTLY CONNECTING THE DATA
CENTERS TOGETHER.IF YOU IMAGINE SOME OF OUR
REGIONS WILL HAVE POTENTIALLY
DOZENS OF DATACENTERS. IF YOU THINK OF HAVING THOSE
DATACENTERS ALL BE ABLE TO TALK
TO EACH OTHER, WELL,YOU COULD
CROSS BAR THEM TOGETHER BUT
THAT’S GOING TO BE A TREMENDOUS AMOUNT OF COMPLEXITY.AND A TREMENDOUS AMOUNT OF CABLE
TO CONNECT THEM ALL TOGETHER.SO WE CREATE REGIONAL NETWORK
GATEWAYS,WHICH ARE REDUNDANT. THERE’S TWO PHYSICAL FACILITIES WHERE THE DATA CENTERS ARE
CONNECTED TO BOTH OF THEM — ALL
OF THEM.NOW TRAFFIC WITHIN A REGION CAN
STAY WITHIN THAT REGION BY GOING
THROUGH THE REGIONAL NETWORK
GATEWAYS BETWEEN ANY OF THOSE
TWODATACENTERS.THOSE REGIONAL NETWORKGATEWAYS
ARECONNECTED TO THE BACKBONE.WHILE THATSIMPLIFIES THE
TOPOLOGY WE WOULD OTHERWISE HAVE
IF WE CONNECT THEM TOGETHER DIRECTLY,MANAGING THEROUTES
ACROSS ALL OF THE PHYSICAL
DEVICES WE’VE GOT ACROSS ALL THE
150 PLUS DATACENTERS THAT WE’VE
GOT OPERATIONAL RIGHT NOW IS
EXTREMELY COMPLICATED.AND THE WAY THAT WE USED TO DO
IT WAS THE WAY THAT TRADITIONAL
I.T.DOES IT.WE’VE HAD SPREADSHEETS AND FILES THAT SHOW THE CONFIGURATION FOR THE ROUTERS AND THE ROUTES.AND THEN WE GO AND WOULD PROGRAM
THE ROUTERS WHEN WE UPDATED OR
EXPANDED CAPACITY, OR CHANGED
OUR ROUTES.WHAT HAPPENS IN THAT CASE,THOUGH — AND HOW MANY OF YOU
ARENETWORK ENGINEERS IN HERE? A FEW OF YOU.HAVE YOU EVER MISROUTED SOMETHING AND BLACK HOLEDTRAFFIC? SO THIS IS ONE OF THE BIGRISKS.THE SYSTEM BECOMES SO
COMPLICATED,NOBODY UNDERSTANDS
ANYBODY THE EFFECT OF A SINGLE
CONFIGURATION CHANGE AND THE
WRONG ONE COULDLITERALLY IMPACT ALL OF AZUREREGIONS AROUND THE
WORLD BY SINK HOLING TRAFFIC.A FEW YEARS AGO,WE STARTED A
PROJECT TO CREATE AN EMULATION
OF OURNETWORK THAT RUNSON
AZURE VIRTUAL MACHINES WHERE THE
PHYSICAL DEVICES ARE REPRESENTED WITH EMULATORS THAT WE CAN GIVETHE EXACT CONFIGURATION WE DID
TO THE REAL DEVICES,TOO,AND
HAVE THEM OPERATE THE SAME WAY
THEY WOULD WITH BUG FOR BUGCOMPATIBILITY.THIS WAY,WE CAN ANYTIME MAKE ACHANGE TO OUR WAN,OUR ROUTES,WE CAN DEPLOY IT TO THIS,RUN
THE SIMULATION AND MAKE SURE
THAT EVERYTHING IS STILL
CONNECTED THE WAY WE WANT TO.AND WE’VE CAUGHT NUMEROUS BUGS
BEFORE THEY’VE GONE INTO
PRODUCTION.THIS ISTHE ONLY WAY WE COULD BE
OPERATING AT THE SCALE WE’RE AT,WITH KEEPING THE NETWORK HIGHLY
AVAILABLE IN THE FACE OF
CONFIGURATION CHANGES.WE SPENTOVER 12 MILLION CORE
HOURS SIMULATING CHANGES TO OUR
NETWORK IN THE LAST YEAR.SO THAT’S A LOOK AT PHYSICAL
NETWORKING.LET’S TALK ABOUT LOGICAL.AND IF YOU TAKE A LOOK AT HOWWE
ARCHITECTED AZURE FROM A LOGICAL
PERSPECTIVE,IT’S USEFUL TO
CONTRAST THAT TO THE TRADITIONAL ENTERPRISE NETWORK ARCHITECTURE WHICH IS BASED OFF OFAPPLIANCES THAT HAVE MANAGEMENT,THE
CONTROL PLANE AND THE DATA PLANE
ALL WITHIN THE SAME BOX.AND WHEN AZURE STARTED, THAT’S
THE WAY WE DID THINGS.WITHIN A YEAR OR TWO AFTER AZURE
LAUNCHED, WE FOUND THIS TO BEHIGHLY COMPLICATED, FRAGILE,EXPENSIVE,ANDLACKED HIGH AVAILABILITY.THOSE BOXES HAVE NO SCALE-OUT
ARCHITECTURE.YOU JUST HAVE TO BUY BIGGER AND
BIGGER BOXES. WHAT WE DID WAS START WORKING ON SOFTWARE-DEFINED NETWORKING
BEFORE IT BECAME A HIPTERM.AND SOFTWARE-DEFINED NETWORKING
YOU TAKE THOSE LAYERS AND BLOW
THEM INTO SOFTWARE PROGRAMS THAT
RUN ON STANDARD SERVERS.SO WE TOOK MANAGEMENT,THAT’S
THE AZURE RESOURCE MANAGEMENT.THAT’S HOW YOU PROGRAM WHAT YOU
WANT FORYOUR VIRTUAL MACHINES,YOUR IP ADDRESSES,YOUR NETWORK
SECURITY GROUPS.THEN WEIMPLEMENTED CONTROLLERS
THAT TAKE THE APIs,TRANSLATE THEM INTO THE GOAL STATE,THE
STATE THAT IS DESIRED FOR THE
COLLECTION OF VIRTUAL MACHINES.AND THEN THE CONTROLLERS TAKE
THAT INFORMATION AND THEN PLUMB
IT DOWN INTO ALL THE DISTRIBUTEDDEVICES,IN THIS CASE SERVERS,THAT IMPLEMENT THE ROUTES AND
THEACLES.THIS ISNETWORKING AT MASSIVE
SCALE.AN EXAMPLE,CREATING A TENANT OR
VIRTUAL NETWORK.THAT NETWORK THEN HASACKLES
PLUMBED INTO IT,GIVES IT TO THE
SERVERS,AND THE KEY TO THIS
FLEXIBILITY IS THATHOST-LEVEL SOFTWARE-DEFINED NETWORKING.HOST-LEVEL SOFTWARE-DEFINED
NETWORKING WILL MAKE SENSE IN
THE CONTEXT OFTHE OVERALL AZURE
ARCHITECTURE.YOU CAN SEE THE PORTAL,COMMAND
LINE INTERFACE, THIRD-PARTY
TOOLS — BY THE WAY, PowerShell IS NOT IN THIS.IT THERE. IT SHOULD BE,BUT I WANTED TO IANNOY JEFFREY,SO I TOOK ITOUT.YOU CAN SEE CONTAINERORCHESTRATORS,COMPUTE NETWORK,THE AZURE FABRIC CONTROLLER WHICH MANAGES THE DEPLOYMENT OF VIRTUAL MACHINES ONTO OUR
SERVERS, AND A HARDWARE MANAGER
RUNNING ON TOP OF THE HARDWARE
INFRASTRUCTURE, RESPONSIBLE FOR
MANAGING THE HEALTH AND THE
PROVISIONING OF THOSE SERVERS. IF WE JUST BLOW OUT THE NETWORK PARTS OF THIS,YOU CAN SEE THAT
THERE IS ANETWORK RESOURCEPROVIDER THAT PLUGS INTO
MANAGER.IT’S RESPONSIBLE FOR
IMPLEMENTINGTHE NETWORK APIs. AND IT’S GOT SIDE-TO-SIDE WITH
THE COMPUTE RESOURCEPROVIDER,THE VIRTUAL MACHINE APIs.I’LL TALK ABOUT THAT IN A LITTLE
BIT.THE NETWORK RESOURCE PROVIDER
HAS COMPONENTS UNDERNEATH IT
THAT MANAGE THOSE CONTROLLERS THAT I TALKED ABOUT IN THE SDM
SLIDE, LIKE THE REGIONAL NETWORK
MANAGER TALKS TO THE NETWORK
STATE MANAGER, THE DIRECTORY
SERVICEKEEPS TRACK OF VIRTUAL
NETWORK PHYSICAL TOCUSTOMER ADDRESS MAPPINGS.AND THEN A SOFTWARE LOAD
BALANCER,WHICH DOES THE LOAD BALANCING.IF YOU TAKE A LOOK AT INSIDE OF
ONE OF OUR SERVERS IT ROUGHLY
LOOKS LIKETHIS.YOU’VE GOT CONTAINERS,VIRTUAL
MACHINES,THAT CAN HAVE
CONTAINERS WITHIN THEM THAT ARE
CONSIDERED FIRST-CLASS OCTOBERS.OBJECTS.THERE’S A NOTE AGENT ON THE HOST
MANAGING THECREATION. AND THEN THERE’S A NUMBER OF
OTHER AGENTSINCLUDING NETWORK
AGENTS LIKE NETWORK MANAGER
AGENT,AND THE LOAD BALANCER
AGENT.THOSE PROGRAM THE VIRTUAL
FILTERING PLATFORM,RESPONSIBLE FORENCAPSULATING TRAFFIC,APPLYING ACKLES AND ROUTES ANDFIREWALL RULES. ANDYOU CAN SEE THAT WHAT WE DO
WHEN WE DEPLOY AVIRTUAL
MACHINE,COMMANDS COME DOWNTO
DEPLOY THE VIRTUAL MACHINE, AND
THE NETWORKING STACK FINDS OUT WHAT IT IS AND TALKS TO ITS
AGENT TO WIRE THAT VIRTUAL
MACHINE UP TO THE NETWORK.SO A NUMBER OF OTHER LOGICALSERVICES THAT WE’RE BUILDING OUT
AS PART OF THE OVERALL LOGICAL NETWORK CAPABILITIES INCLUDE
SOMETHING CALLED THE AZURE FRONT
DOOR SERVICE, WHICH WE ANNOUNCED
HERE AT IGNITE. I TALKED ABOUT THOSE 130 POINTS OF PRESENCE,AND WE ARE
LAUNCHING AZURE FRONT DOOR
SERVICE IN ABOUT 70 OF THOSE.AND WITH THIS,YOU CAN TERMINATE
SSL TRAFFIC AT OUR FRONT DOORS.YOU CAN DO GLOBAL HDP LOAD BALANCINGTHROUGH OUR FRONT
DOORS. IT’S INTEGRATED WITH AZURE WEBAPPS, IT’S USING THAT FOR
APPLICATION ACCELERATION OF OUR
FRONT DOORS AS WELL.AND WE’VE BEEN USING AZURE FRONT
DOOR INSIDE MICROSOFT ACROSS OUR
PROPERTIESLIKE DYNAMICS 365 AND
OFFICE 365 AND BING,AND XBOX LIVE FOR THE LAST FIVE YEARS.SO THIS IS US BRINGING THAT
INTERNAL SERVICE AND MAKING IT
AVAILABLE FOR CUSTOMERS TO USE.ANOTHER SERVICE THAT WEANNOUNCED THIS WEEK IS SOMETHING CALLEDAZURE VIRTUAL WAN.THIS IS GOING AFTER THE PROBLEM
WHERE YOU’VE GOT LOTS OF
DIFFERENT USERS RUNNING AROUND,SOME OF THEM THAT ARE LOCATED NEAR YOUR CORPORATEDATACENTER,SOME THAT ARE BRANCHOFFICES,SOME ARE USING MOBILEDEVICES,AND YOU WANT CONTROL OVER THEIR
TRAFFIC.YOU WANT TO BE ABLE TO FILTER
THEIR TRAFFIC AND DO SECURITY
INSPECTION.BUT GETTING A HANDLE ON IT IS
HARD.YOU COULD MAKE THEM ALLVPN,SEND ALL THE TRACK BACK INTO
YOUR CORPORATE NETWORK, BUT THEN
YOU’RE OVERLOADING THENETWORK, YOUR FIREWALL DEVICES GET
OVERLOADED. SO WHAT WE’RE DOING IS MAKINGIT SO YOU CAN VERY EASILY PLUG YOUR BRANCH OFFICES AND YOUR VPNGATEWAYS INTO AZURE,INTO A
BACKBONE,THE VIRTUAL WAN
BACKBONE THROUGH HUBS SO THAT ALL THE TRAFFIC GETS ROUTED TO
THE RIGHT PLACES IN THE CLOUD.AND RELATED TO THAT IS THE
GENERAL AVAILABILITY WE
ANNOUNCED HERE OF THEFIREWALL SERVICE.IN THIS CASE,WHAT’S BEHIND THE
SCENES IS YOU CREATE A VIRTUAL FIREWALL.IT’SLAUNCHING VIRTUAL MACHINES THAT ARESERVERLESS.WE’RE SCALING THEM OUT BASED ON
YOUR TRAFFIC LOAD, THE LEVEL OF
SERVICE THAT YOUREQUESTED. AND THIS FIREWALL CAN IMPLEMENT LEVEL TWO ALL THE WAY UP TO
LEVEL SEVEN POLICIES.SO YOU CAN USEQDN KIND OF RULES
TO BLOCK TRAFFIC,OR ROUTE
TRAFFIC.YOUCAN USE IP ACKLES AS WELL ON
THE FIREWALL.YOU CAN HAVE THEFIREWALL
ALLOCATE THE IP ADDRESS FROM A
SPECIFIC SUBNET,MAKING IT VERY
EASY TOSET UP FIREWALL RULES THAT WILL SCALE WITH CLOUDDEPLOYMENTS ON CORPORATEFIREWALLS.SO I MENTIONED THAT VIRTUAL
MACHINES AREFIRST-CLASS CITIZENS IN AZURE NETWORKING.BUT UP TO THISPOINT,CONTAINERS
HAVE NOTBEEN.WE ARE WORKING ON A PROJECT HERE
CALLED SWIFT WHICH WE’RE ROLLING OUT RIGHT NOW FOR OUR OWN
SERVICES AND WILL MAKE AVAILABLE
TO THIRD PARTIES THAT LETS AN
ORCHESTRATOR PROGRAM THEAZURE
VIRTUAL NETWORK DIRECTLY.YOU HAVE CONTAINERS THAT BELONG
TO DIFFERENT APPLICATIONS THAT
SHOULDBE IN DIFFERENT VIRTUAL NETWORKS THAT YOU’RE DEPLOYING
TO A CLUSTER.HOW DO YOUDO THAT?UP UNTIL NOW,VIRTUAL MACHINES
ARE IN A SINGLE VIRTUAL NETWORK.WITH SWIFT,YOU CAN DELEGATE
MANAGEMENT OFA SUBNET TO AN
STATER AND THE VIRTUALMACHINES. THE ORCHESTRATOR CAN CALL INTO
THE AZURE NETWORK AGENT ON THESERVER AND ASSIGN IP ADDRESSES
TO PARTICULAR VMs FROM THAT
SUBNET RANGE.WITHIN THE VM IT WILL RUN A
CONTAINER NETWORK PLUG-IN THAT
TAKES THAT IP ADDRESS AND
EXPOSES IT TO THE CONTAINER, OR
MAPS IT TO WHAT THE CONTAINER SEES.SO IN THIS WAY,AS THECONTAINER MOVES FROM VIRTUAL MACHINE TO
VIRTUAL MACHINE,THEORCHESTRATOR KNOWS WHICH VIRTUAL
MACHINE IT’S ON AND CAN TELL
USING AN AGENT,CAN TELL THE AGENT ON THE HOST WHICH SUBNET,WHICH IP ADDRESSES FROMTHE
SUBNET IT’S MANAGING,THE
CONTROLLER IS ON THE MASTER
NODES AND WILLTELL THE AGENTS THAT ARE DEPLOYING CONTAINERS
WHICH IPADDRESSES THEY SHOULD
PLUMB INTO THE HOST TO SAY HEY,THIS CONTAINER IS HERE,THE IP
ADDRESS SHOULD BE THIS.AND IT MIGHT BE KIND OF HARD TO
FOLLOW,SO LET’S GO AHEAD AND TAKE A LOOK AT A DEMO OF THAT IN ACTION.SO WHAT I’M GOING TO DO HERE ON
THEDEMO MACHINE IS I’VE GOT A KUBERNETESCLUSTER AND I’M GOING TO SHOW YOUTHAT THAT CLUSTER
CONSISTS OFTWO NODES,A WORKER
ANDA MASTER. SO,HERE’S THE WORKER.AND WHAT I’M GOING TO DO FOR
THIS WORKER NOWIS DEPLOY SOME
PODS TO IT.BUT LET’S FIRST SHOW THAT
THERE’S NO PODS RUNNINGON THIS.GET PODS.NO PODS RUNNING.SO I’M GOING TO RUN A LITTLE SCRIPT.WHAT THIS SCRIPT IS DOING IS DEL
DELEGATING ASUBNET TO THE KUBERNETES MASTER AND SAYING YOU CANUSE THIS TO ASSIGN THEM TO
WORKERS ON THAT WORKER VM.WE’RE ASSIGNING TWO VIRTUALNETWORKS TOKUBERNETES. NOW WE CAN START TO DEPLOY
CONTAINERS INTO ONE VIRTUAL
NETWORK OR THE OTHER ONE,AND
THEY’RE ISOLATED IN THE SAME WAY
TWO VIRTUAL MACHINES WOULD BEISOLATED,EVEN THOUGH THOSE CONTAINERS HAPPEN TO BE RUNNINGON THE SAME VIRTUAL MACHINE.AND TO DEMONSTRATE THAT IN
ACTION,LET’S SEE IF THIS THING
HAS COME UP WITH IP ADDRESSES.AND THEY BOTH HAVE THE SAME IP
ADDRESS,EVEN THOUGH THEY’RE
BOTH RUNNING — AND THESE ARE
REAL,AZURE IP ADDRESSES THAT AREASSIGNED TO VIRTUAL
NETWORKS, SO THEY’REACCESSIBLE
FROM OTHER VIRTUAL MACHINES
TALKING TO THOSE IP ADDRESSES.AND SO WHAT I’M GOING TO DO IS
LAUNCH A WEB SERVER ON ONE OF
THEM.SO I’MGOING TOACTUALLY CUBECONTROL,EXEC DASH I.T.INTO
THIS FIRSTCONTAINER.WHOOPS.COPY,PASTE, BASH.AND TO SHOW YOU THAT THIS IS NOT
A CANNEDDEMO,WHAT MESSAGE DO YOU WANT ME TO WRITE?HELLO AZURE. LET ME BE CORRECT HERE.HELLO,AZURE. OH,HELLO,CAPITAL AZURE.[ LAUGHING ]
>>Mark: OKAY.NOW LET’S START UP A WEBSERVER.23.0.15,I THINK THAT’S IT.LET’S PUT IT ON PORT 80.NOW THAT WEB SERVER IS RUNNING
INSIDE ONE OF THESE CONTAINERS
THAT’S MAPPED TO ONE OF THOSE.NOW,I’M GOING TO TAKE FROM
ANOTHER VIRTUAL MACHINE THAT’S
OUTSIDE A DIFFERENT NODE FROMTHEKUBERNETES WORKER THAT’S PARTOF THE SAME VIRTUAL NETWORK THAT WEB SERVERIS IN.SEE IF I CAN GET THAT FILE.AND THERE IT GOES. NOW,THIS OTHER VIRTUAL MACHINE
IS IN THATSECOND SUBNET,THE
ONE WITH THE CONTAINER THATDOESN’THAVE THE WEB SERVER
RUNNINGIT.AND YOU CAN SEE THATI CAN PING
THAT CONTAINER, THE SECOND
CONTAINER.AND TO PROVE THAT IT’S ACTUALLY
A DIFFERENT CONTAINER EVEN
THOUGH IT’S GOT THE SAME IP ADDRESS ANDA DIFFERENT VIRTUALNETWORK, AND THERE.IT’S NOT RUNNING THE WEB SERVER.SO WHAT WE’VE JUST DONE IS SHOWN
TWO CONTAINERS WITH THE SAME IP
ADDRESSEXPOSED FROMTWO
DIFFERENT VIRTUAL NETWORKS ON
THE SAME MACHINE. THIS ALLOWS CONTAINERS TO HAVE
THE SAME NETWORK POLICIES APPLY
TO THEM AS VIRTUAL MACHINES DO,FIRST-CLASS CITIZENS IN THE
AZURE NETWORK.SO LET’S TAKE A LOOK INSIDE OF
OUR AZURE SERVERS NOW.I SAID I WAS GOING TO TAKE YOU
ON A WALK DOWN MEMORY LANE.STARTING WITH OUR GEN TWO
MACHINE, I’VE GOT RATIOS OF CORE
AND RAM TO GIVE YOUAN IDEA OF
HOW WE’VE GROWN.IN THE ORIGINAL DAYS,WE STARTED BYSCALE-OUT COMMODITY SERVERS.32 GIG OFRAM,ONE GIG BITNIC,NO SSD, IT WASAN AWESOME
MACHINE BACKIN 2010.AND THEN WE INTRODUCED OUR THREE MACHINES,LIGHTLYLARGER,TEN ANDTHEN 40 GIGABITNETWORKING,SSDs. THEN WE INTRODUCED A SCALE-UP
MACHINE. WE CALLED THIS ONEGODZILLA
BECAUSE AT THE TIME,IT WAS ONE
OF THE LARGEST SERVERS IN THE
PUBLICCLOUD.ITHAD 512 GIGABYTESOF RAM.AND THIS WAS TORUN SCALE-UP SAP
WORKLOADS AND MEMORYDATABASES.WE HAD LOTS OF PLACES MIGRATINGHANNAH.THEN WEINTRODUCED OUR GEN FIVE
GENERAL PURPOSE MACHINE. IT’S HALF THE MEMORY NOW,OR
MORE THAN HALF THE MEMORYOF OUR
GODZILLA, SO STARTING TO GET
BIGGER GENERAL PURPOSE MACHINES.HERE’S OUR GEN SIX,IN THE
PROCESS OF DEPLOYING NOW.BUT THEN WE HAD CUSTOMERS SAYING WE’VE GOT EVEN BIGGER SAP HANNAH WORKLOADS,THEY DON’T FIT IN
YOUR GODZILLA,WHAT ARE YOU
GOING TO DO FOR US?SO WE INTRODUCED THIS.[ LAUGHING ]
>>Mark: AND THIS IS CALLED
THE BEAST.IT’SGOT FOUR TERABYTESOF RAM. YOU’RE IMPRESSED WITH THAT.WELL, THEY CAME AND SAID,OKAY,THAT’S GOOD BUT WE’VE GOT EVEN
BIGGERS.A.P.WORKLOADS. SO WE JUST ANNOUNCED THIS,BEASTV2. 12 TERABYTES OF RAM.YOU MIGHT ASK WHY WE DIDN’T CALL
IT SON OFBEAST. [ LAUGHING ]>>Mark: I GUESS.OR BUTTE BEAUTY AND THE BEAST.IT TURNS OUT THERE’S SOME KIND
OF MOVIE THAT HASTHAT NAME.BUT LET ME SHOW YOU BEASTV2.OH,BY THE WAY WE ALSO HAVE SPHERE.THAT’S ANOTHERONE.[ LAUGHING ][ APPLAUSE ]>>Mark: OKAY.SOHERE’S BEAST.SO,I’M GOING TO SHOW YOU ONE OF
MY FAVORITE TOOLS.[ LAUGHING ]>>Mark: AND SHOW YOU JUST HOW MUCH RAM THIS THING HAS IN IT,JUST FOR REAL.ALL RIGHT.WHOEVER WROTE THIS,IT’S LIKE
CUT OFF.IT’S GOT A NEGATIVE NUMBER HERE.[ LAUGHING ]>>Mark: THIS IS A SAD,SAD
DAY. I’VE GOT TO RUN THISTOOL
INSTEAD. [ LAUGHING ][ APPLAUSE ]>>Mark: SO,HOW MUCH RAM DOES
THIS HAVE IN IT?THERE IT IS RIGHT THERE.SO,NOTEPAD IS AMAZINGLY FAST ON
THIS THING. [ LAUGHING ]>>Mark: AND YOU CAN ACTUALLYOPEN SEVEN CHROME TABS IN IT
WITHOUT ANY PROBLEM.[ LAUGHING ]
[ APPLAUSE ]>>Mark: OKAY.SO,NOW, WE’VE ALSO GOT SPECIAL
PURPOSE SERVERS. HERE’S HPC SKEW.WE STARTED TO INTRODUCING GPUs,SO ONE OF OUR FIRST
SERIES HAD A K80 PROCESSOR FOR
GENERAL PURPOSE COMPUTE ONGPUs,WE INTRODUCED AHIGH-PERFORMANCE COMPUTING VMSWITH INFINIBINNETWORKING.WE INTRODUCED HIGH-PERFORMANCE
COMPUTING SKEWS THAT ARE COMPUTE
INTENSIVE SO THEY RUN AT HIGHER
FREQUENCY AND HAVE MORE CORES.ANDTHEN WESTARTED TO INTRODUCE
DEEP LEARNING SKEWS.THE NDV2,THISONE HAS EIGHTNVIDIAV100S LINKED TOGETHER INA FOUR-U SERVER BOX.THOSE THINGS CONSUME QUITE A BIT
OF ELECTRICITY.HERE’STHELV2. THIS IS A HIGH-DENSITYSSD SKEW
FOR DATABASE-TYPEAPPLICATIONS,IT STORES DATALOCALLY.WE PROVIDE HIGH-PERFORMANCE WITH
THISSKU USING DIRECT.SO WEEXPOSE THESSDs
DIRECTLY,THEIOQs DIRECTLY INTO THE VIRTUAL MACHINES,MAP
THEM DIRECTLY SO THEY CAN
DIRECTLY ACCESS THEIOQs,PUTTING COMMANDS INTO IT AND
READING THE RESULTS OUT OF THEM. BUT WE PUT FILTERS ON THE
COMMAND QUEUE SO THAT THE VM
CANNOT DO SOMETHING LIKE FLASH
THE FIRMWAREOR OTHER THINGS
THAT IT SHOULDN’T BEDOING TO
THE SSD.THIS PROVIDES SECURITY WHILE
GIVING THEVM NATIVE SSD
PERFORMANCE. LET’S SEE WHAT KIND OF
PERFORMANCE WE CAN GET FROM
ONEOF THESE THINGS. SO THIS TAB IS ANNVME DIRECT
VM. WE’RE ACTUALLY LOOKING ATA
AZURE SERVER HOST WITH TASK
MANAGER RUNNING ON THE HOST.THIS IS THE VM WITH AN
APPLICATION RUNNING THAT I’M GOING TO LAUNCH.YOU CAN SEEPERFMON HEREAND
YOU CAN SEE THATWE’VE GOT 0IOPs GOING RIGHT NOW.I’M GOING TO PRESS ENTER TO RUN
THIS SCRIPT, WHICH IS GOING TO
START TO HAMMER THE DISK.AND THIS IS GOINGTO ACQUIESCE ATABOUT 3.7 MILLION IOPs OFFOF ONE SERVER.THERE’S A BUNCH OF DISKS IN
HERE.SO THIS IS ACTUALLY CLOUD LEADING RIGHT NOW.THERE IS NO OTHER CLOUD SERVER
THAT DELIVERS IOPs THIS HIGH.YOU CAN SEE EVIDENCE IT’S MAPPEDDIRECTLY,NO ACTIVITY ON THE HOST,HERE’S INSIDE THE VM,IT’S
BRINGING A LOT OF CORES TO DO
THAT.SO THAT’S A LOOK ATNVME.SO GOING TO AZURE COMPUTE NOW,LET’S TAKE A LOOK AT THE
ARCHITECTURE AGAIN,BUT THIS TIME WE’RE GOING TO FOCUS IN ON THECOMPUTESTACK.AND HERE YOU CAN SEE THE
DIFFERENT LAYERS IN THE COMPUTE
STACK.YOU CAN SEE THE GLOBAL RESOURCE
MANAGER,RESOURCE PROVIDERS THAT
OPERATE AT THE REGIONAL LEVEL, I’VE TALKED ABOUT THE NETWORKING RESOURCE PROVIDER,WE’RE
FOCUSING ON THECOMPUTE RESOURCE
PROVIDER HERE.YOU CAN SEE THE REGIONAL
CONTROLLERS, REGIONAL DIRECTORYSERVICE,THEUSLB SERVICE,ALL
OF THOSE ARE GOING TO BE
COORDINATING WITH CLUSTER-LEVEL SERVICES FOR SMALLER BLAST
RADIUS MANAGING A SMALL SUBSET
OF THE SERVERS AND THEN TALKING
TO THE NODES.AND WHEN YOUDEPLOY A VIRTUALMACHINE, THIS IS THE FLOW,BECAUSE THEY’RE ALL
ORCHESTRATING TO GET THAT DATA DOWN TO THE SERVERS SO THAT YOU GET YOUR VIRTUAL MACHINE UP WITH A DISK ATTACHED TO IT AND
NETWORKING. SO WE’RE GOING TO SPEND THE NEXT TEN MINUTES OR SO GOING THROUGH THIS.THAT’S JUST A JOKE,WE’RE NOT GOING TO DO THAT UNLESS YOU WANT TO.ALL RIGHT.WE’LL SKIP THAT.BUT WE’VE BEEN DOING A BUNCH OF
THINGS TO MAKE VIRTUAL MACHINES
EASIER TO MANAGER.ONE OF THE THINGS WE’RE EXCITED
TO ANNOUNCE IS GIVING SERIAL CONSOLE ACCESS TO YOUR VIRTUAL
MACHINES. SO IF YOU’RE NOTIMPRESSED —
[ APPLAUSE ]
>>Mark: HOLD YOUR APPLAUSE.I’LL SHOW YOU THIS IN ACTION.>>Mark: OKAY.SOWE GOTO . . .BACK HERE AND GO TOTHE BROWSER.I’MGOING TO REFRESH THESE AND
WE’RE GOING TO COME BACK AND
I’LL SHOW YOU WHAT THOSE ARE IN
A SECOND.BUT WHAT I WANT TOSHOW YOU HEREIS I’VE GOTA LINUX VIRTUAL MACHINE HERE RUNNING IN AZURE.AND YOU CAN SEE THAT IT’S GOT
TWO NETWORK INTERFACES ON IT.ONE OF THEM IS THE ETHERNET
NETWORK INTERFACE CONNECTED TO
THE AZURENETWORK.AND I’M GOING TO DO SOMETHING
THAT’S GOING TO SIMULATE A
PROBLEM THAT WILL COME USUALLY
BECAUSE OF SOME CORRUPTION, SOMEBUG.BUTSHOWS YOU . . .ACTUALLY, I’M GLAD I DIDN’T DO THAT,BECAUSE I WANTED TO PING THIS
VIRTUAL MACHINE FIRST TO SHOW
YOU.SO I’M PINGING THAT VIRTUAL
MACHINE.I’LL JUST TYPE THAT.AND WHAT THIS COMMAND DOES IS
TAKES OUT THATETHERNET
INTERFACE. SO THE SECOND I DOTHIS,THOSE
PINGS STOP.AND AT THIS POINT YOUR VM IS COMPLETELY INACCESSIBLE.ORIS IT? IF WE GO BACK TO THE SERIALCONSOLE, I’M SSH’D INTO MY VM
THROUGH THE SERIALPORT ON THE
VIRTUAL MACHINE BY LEVERAGING AN
AGENT IN THE AZURE HOST,THE
SERVER.SO,IF I DOTHIS,BOOM. I’VE JUST FIXED THE NETWORK.[ APPLAUSE ]>>Mark: BUT LET’S TAKE A
LOOK. THAT’S NOT JUST FOR LINUX,THAT’S FOR WINDOWS,TOO.I’M GOING TO CRASH WINDOWS HERE
USINGANOTHER ONE OF MY FAVORITE
TOOLS CALLED NOT MY FAULT.I WROTEIT FOR WINDOWS
INTERNALS. AND YOU CAN CRASH THE SYSTEM IN LOTS OF COOL WAYS.I’M GOING TO CRASH IT IN A
SIMPLE WAY.I’VE JUST PRESSED THE CRASH
BUTTON AND THIS THING IS GONE.SO NOW WE WANT TO SEE WHAT HAPPENED TO THIS.IF WE GO TO THE CONSOLE FROM MY
WINDOWS MACHINE YOU CAN SEE WHATHAPPENED.IT CRASHED. MY FAULT GOT THIS INFORMATION
SPIT OUT AT THE CRASH THROUGH
THE SERIAL PORT.AND NOW WE CAN SEE IT.AND ALSO, BECAUSE WE’RECAPTURING SCREENSHOTS OF THE VM
INTO AZURE STORAGE WE CAN SHOW
YOU THAT IN THE PORTAL, TOO.SO YOU CAN SEE HERE WHAT
HAPPENED WITH THAT VM.SO THAT’S A LOOK AT SERIAL
CONSOLE ACCESS.SO,ANOTHER WAY THAT WE’RE
ADVANCING THE STATE OF ART WITH
CONTAINERS BESIDES GIVING THEM
FIRST-CLASS CITIZENSHIP IN THE
AZURE NETWORK IS TO TAKE THEM SERVERLESS.THIS IS SOMETHING WE ANNOUNCED
HERE.YOU MIGHT HAVE SEEN SCOTT OR
CORY TALK ABOUT IT.IT’S VIRTUALCUBELET.THIS IS EFFECTIVELY AN AGENT
THAT PLUGS INTO A KUBERNETES CLUSTER AND BEHIND IT IS
WHATEVER IT WANTS.COULD BE A CLUSTER OF ANOTHER TYPE OF COMPUTE.BUT TOKUBERNETES IT JUST
PRESENTS ITSELF ASA NODE, ANOTHERSERVER WITH INFINITE
CAPACITY. WE CREATED ONE THAT CONNECTS
WITH THESERVERLESS CONTAINER
SERVICE. YOU CAN LAUNCHA DOCKER
CONTAINER INTO ACI.THE CONTAINER WILL JUST LAUNCH
INTO A SERVERLESS
INFRASTRUCTURE. AND SO VIRTUAL CUBELET ON TOP OF ACI EFFECTIVELY MAKES KUBERNETES SERVERLESS.IF YOU HAD A CLUSTER WITH ZERO
NODES EXCEPT A VIRTUAL CUBELET,YOU COULD START DEPLOYING PODS
TO IT AND THEY WOULD
AUTOMATICALLY SCALE OUT ON TOACI.WE’RE ALSO TAKING SERVICEFABRICSERVERLESS.HOW MANY PEOPLE HAVE HEARD OF
SERVICE FABRIC? IT’S OUR OWNMICROSERVICES
ORCHESTRATOR. IT SUPPORTS HIGHER-LEVEL
PROGRAMMING LANGUAGES,STATEFULMICROSERVICES, WE BUILT A
TREMENDOUS AMOUNT OF AZURE ON
SERVICE FABRIC.YOU CAN SEE THESERVICES,LIKE
THOSE CORE AZURE RESOURCE
PROVIDERS ARE ALL BUILT ON
SERVICE FABRIC.CORTANA, COSMOS DB IS BUILT ON
SERVICE FABRIC, SKYPE FORBUSINESS,EVENT HUBS,EVENT ,THE LIST
GOES ON AND ON,ALL OF THESE
OPERATING AT MASSIVE SCALE FOR
YEARS.WE MADE THIS AVAILABLE OPEN
SOURCE.WE MADE IT AS A SERVICE IN AZURE
SO YOU CAN LAUNCH A SERVICE FABRIC CLUSTER.BUT WE’RETAKING IT SERVERLESS
WITH A PREVIEW OF MESH. WITH MESH,WE MANAGE SERVICE
FABRIC CLUSTERS IN THEBACKGROUND.YOU LAUNCH SERVICE FABRIC APPS
TO THE MESH SERVICE.YOUDON’TSPECIFY VIRTUAL MACHINES.YOU JUST SPECIFY THEM AS CONTAINERIZED MICROSERVICES AND THEN WE LAUNCH THEM AND MANAGE
THEIR VIRTUAL MACHINES
UNDERNEATH THEM AND DO ALL THE
WIRING,LOAD BALANCING, AND
EVERYTHING ELSE THAT’S NECESSARY TO KEEP THEM RUNNING.I’M GOING TO SHOW YOU A QUICK
DEMO OF SERVICE FABRIC MESH NOW
AND SHOW YOU KIND OF THE SCALE
CAPABILITIES THAT YOU CAN GETWITH THIS.ACTUALLY,THAT’S A DIFFERENT
DEMO.YOU’RE FREE TO STAND OUT HERE IF
YOU WANT.NO,I’D JUST KIDDING.IT’S GOING TO BE THE NEXTDEMO.SORRY.SO HERE I’VE GOT A MESH
APPLICATION THAT CONSISTS OF A
WEB FRONT END AND THREE
MICROSERVICES THAT ARE THE SAME.THEY’RE FIREWORKS MICROSERVICES
AND WE PASS IN — THEY EACH HAVE
A REPLICA COUNT OF ONE.AND WEPASS IN THE — HERE’S ONEOF THEM.REPLICA COUNTOF ONE.YOU CAN SEE WE PASS IN WHICH
COLOR THE FIREWORK SHOULD BE AS
ENVIRONMENT VARIABLES.THAT’S THE RED ONE.IF WE SCROLL DOWN,WESEE THE
GREEN ONE, REPLICA COUNT OF ONE
AND FINALLY A BLUE ONE.AND WHAT THIS IS DOING IS THIS
RIGHT HERE.THAT’S WHAT THIS APPIS DOING.JUST FIRING A FEW EVERY SECOND.RED,BLUE,GREEN,KIND OF
RANDOMLY. WE WANT TO GET SOMETHINGA
LITTLE MORE EXCITING,SO WHAT
WE’RE GOING TO DO IS LAUNCH THIS
VERSION OF IT WHERE ALL WE DID
WAS TAKE THAT APPLICATION,SETTHE REPLICA COUNT TO 500 AND I’M
GOING TO LAUNCHIT RIGHTNOW AS
AN UPDATE TO THE EXISTING ONE
RUNNING ON SERVICE FABRIC MESH. IN ABOUT SOMEWHERE AROUND 30
SECONDS OR SO,WE’RE GOING TO
START TO SEE THIS THINGSCALEOUT.>>Mark: SHOULD BECOMING
ALONG.THEREWE GO.WE’RE SCALING OUTREPLICAS.[ APPLAUSE ]>>Mark: SO WE DIDN’T HAVE TO
SAY HOW MANY VIRTUAL MACHINES WE
WANT. WE’RE ONLY PAYING FOR THOSE
CONTAINERS AS THEY RUN.IF WE SHUT THEM DOWN, WE’RE NOT
PAYING ANYTHING MORE FOR ANY
COMPUTE,MEMORY,OR ANYTHING
ELSE. SO THIS IS THE POWEROF
SERVERLESS RIGHTTHERE.SO THE OTHER THINGWE’RE DOING
WITH CONTAINERS IS TAKING IT TO
THE EDGE.AND MICROSERVICES IN GENERAL.SO YOU’VE HEARD ABOUT INTELLIGENT CLOUD,INTELLIGENT
EDGE, OUR VISION FOR COMPUTING
THAT SPANS BOTH.AND WE WANT THE COMPUTING TO BE
CONSISTENT ACROSS BOTH.SO THAT SERVICE FABRIC
APPLICATION,WE WANT TO MAKE IT
POSSIBLE FOR YOU TODEPLOY THATONRASPBERRYPI CLASS DEVICES. WE WANT YOU TO BE ABLE TORUN IT ON DEVICES IN THE REAL WORLD.SAME SET OFTOOLING.ORCHESTRATION OF THE LAUNCHING
AND UPDATING OF THOSE APPLICATIONS THROUGH A
CONSISTENT CLOUD EXPERIENCE.SO WHAT I’M GOING TO SHOW YOU IS
DEPLOYING A HIGHLY AVAILABLE
CLUSTER TO A GROUP OF EDGE
DEVICES.YOU CAN COME OUTNOW. [ LAUGHING ]
>>Mark: I’VE GOTTHIS SETUPUPHERE.NOW,THIS IS A VERY EXPENSIVE PIECE OFMACHINERY.THIS IS EXTREMELY EXPENSIVE ANDVERY DANGEROUS.IT’S GOT A SPINNING WHEELONIT
THAT IF I PUT MY HAND IN IT,I
WOULD GET IN TROUBLE.YOU CAN SEE THAT IT’S GOT A
SENSOR.THAT IS PREVENTING ME FROM
HURTING MYSELF.NOW,IF I GO TO THE DEMO MACHINE
I’LL SHOW YOU THAT I’VEGOT,ACTUALLY,THAT APP RUNNING UP
HERE IN THE CLOUD.AND THIS SHOULD BESHOWING ME
THAT TELEMETRY COMING FROM THAT
THING.IT’S NOT. WHAT WE SHOULD BE SEEING HEREISDATA SPEWING OUT HERE FROM THAT
DEVICE IN THIS CLOUD-BASEDPORTAL WITH THESE APPLICATION
AND MACHINE LEARNING ALGORITHMSALL RUNNING HERE IN THE CLOUD.I’M GOING TO PUSH THATAPPLICATION DOWN TO THEDEVICES.SO IF I GO TO THE IOTDEVICES
HERE,I’M GOING TO CREATE THIS
HIGHLY AVAILABLE CLUSTER OUT OF
THE COLLECTION OF THOSE THREE DEVICES.AND FOR SOME REASON I’MNOT SEEING THOSE DEVICES.HMM.IT SEEMS LIKE I’VE GOT A
NETWORKINGPROBLEM, APPARENTLY.THAT’S PREVENTING THIS DEMO FROM
WORKING.AND IT IS CONNECTED TO THE
NETWORK.SO I SHOULD BE SEEING THE
DEVICES HERE,BUT I’M NOT FOR SOME REASON.SO — HERE WE GO,SORRY. THEY’RE RIGHT HERE.THERE THEY ARE. I’M GOING TO SELECT THEM ALL AND SAY CREATE EDGE CLUSTER.I WAS LOOKING IN THE WRONG
PLACE.I’M GOING TO CALL THIS IGNITE
EDGE CLUSTER.THIS IS GOING TO CREATE A HIGHLY
AVAILABLE CLUSTER YOU CAN SEE
THIS,IT’S PROVISIONING. IT’S DEPLOYING A CLUSTER TOTHE
THREE IOT DEVICES THAT ARE
CONNECTED TO THAT VERY EXPENSIVE
PIECE OF MACHINERY. NOW I’M GOING TO DEPLOY A BUNCH OF APPS TO THEM.SO RUNON WEB APP, RUNINIGNITE,THIS EDGE.SOI’M DEPLOYING THE MACHINE
LEARNING PROTECTION APP USING
VISION RECOGNITION TO DETERMINE
IF THERE’S SOME OBJECT EXTRACTING THOSE IN FRONT OF
THAT DEVICE,AND IT’S GOING TOTURN IT OFF.SO THIS WILL FINISH DEPLOYING ON
THOSE DEVICES HOPEFULLY IN A SECOND HERE.AND THE WHOLE IDEA HERE IS I
WANT THIS THING TO BE NOT
AVAILABLE BECAUSE THIS MACHINE
CAN’T BE OFFLINE BUT IT NEEDS TO
BE SAFE.I NEED TO RUN THE MACHINE
LEARNING ALGORITHM ON THAT
DEVICE AND HAVE IT BE ABLE TO
TOLERATE FAILURES OF COMPONENTS LIKE THOSERASPBERRYPIS.THIS MIGHT BELAGGING.DO A REFRESH HERE.AND YEAH,OKAY,THEY’RERUNNING.THEIR STATUS ISRUNNING.WHAT I’M GOING TO SHOW TO
DEMONSTRATE THIS AVAILABILITY IS
I’M GOINGTO UNPLUG THIS DEVICE AND THE NETWORK HUB THIS DEVICE ISCONNECTED TO FROM THE
NETWORK. SO I’VE JUST UNPLUGGED IT RIGHT THERE.AND NOT ONLY THAT,BUT IF — ONCE I’VE DONE THAT,IF I COME
BACK OVER TO THE MACHINE,I’VE
GOT ANOTHER VIEW OF THE PORTAL
RUNNING ON THIS SYSTEM THAT IS
TALKING TO THOSEDEVICES. ANDYOU CAN SEE THAT IT STOPS
COLLECTING TELEMETRY.I DON’T KNOW WHAT JUST HAPPENED.THAT’S NOT WHAT I MEANT TO DO.THAT’S THE WRONG — HERE WE GO.SO,YOU CAN SEE THAT THE MACHINE LEARNING APP IS RUNNING ON EDGE DEVICE NUMBER 3.AND SO WHAT I’M GOING TO DO IS
GO UNPLUG EDGE DEVICE NUMBER 3.AND AT THIS POINT WESHOULD SEETHAT THIS MONITOR,SEE THAT THAT DEVICE IS OFFLINE.WE’RE GOING TO SEE AN AUTOMATIC
THE CLUSTER IS DOING,BECAUSE I HAVE BOTH KUBERNETES AND SERVICE FABRIC APPS ON THIS THING.YOU CAN SEE THAT I’MSTILLSAFE. SO THAT ISDISCONNECTED,HIGHLY
AVAILABLE DEPLOYMENT FROM THE
CLOUD USING MICROSERVICES THAT
ARE ALSO CONSISTENT ACROSS THE
CLOUD AND ON-PREM.SO THAT’S WHERE WE’RE TAKING
THAT.THANKS. [ APPLAUSE ]>>Mark: NOW,THE OTHER THING — ONE OF THE OTHER THINGS WE’RE DOING WITH COMPUTE IS THAT WE WANT TO PROTECT YOUR DATA.IF YOU CAME TO MY SESSION
EARLIER IN MECHANICS YOU SAW ME
TALK ABOUT CONFIDENTIAL
COMPUTING.THE IDEA IS THERE’S LOTS OF
THREATS THROUGH YOUR DATA ANDPOTENTIALLY DIFFERENT KINDS OF
THREATS WHEN YOU PUT YOUR DATA
IN THE CLOUD.THERE’S THIRD PARTIES THAT MIGHT
REQUESTACCESS, THERE’S HACKERS
THAT MIGHT BREACH THE
INFRASTRUCTURE AND GET ACCESS. AND SO WE TAKE LOTS OF STEPS IN DEFENSE.WE SPEND A BILLION DOLLARS IN
CYBERSECURITY.BUT WE ALSOIMPLEMENT CAPABILITIES LIKE ENCRYPTION FOR OUR SERVICES AT REST,ENCRYPTION IN TRANSIT,APPLICATIONS CAN
ENCRYPT THEIR OWN DATA IN
TRANSIT,BUT WHAT’S BEING
MISSING IS PROTECTING THE DATA
WHILE IT’S IN USE.THE CLOUD IS NOT VERY USEFUL
UNLESS YOU CAN PROCESS THAT DATA
TO GET INSIGHTS, MACHINE
LEARNING,ANALYTICS,AND IF YOU
CAN’T DO THAT WHILE FULLY
PROTECTING IT,THEN YOU’RE
MISSING ONE OF THESE THREE.SOWITH CONFIDENTIAL COMPUTING, BASED OFF THE BLACK BOX
TECHNOLOGY CALLED TRUSTED
EXECUTION ENVIRONMENTS, YOU CAN
STICK YOUR DATA IN A TDE AND
IT’SISOLATED FROM EVERYTHING ELSE.THERE’S TWO TYPES,SOFTWARE AND
HARDWARE.A HARDWARE ONE IS INTEL, THE
PROCESSOR CREATES THE BLACKBOXES, ENCRYPTED PORTIONS OF RAM
WHERE THE CPU IS IN A SPECIAL
MODE AND NOTHING CAN GET IN
EXCEPT FORTHE CODE INTENDED TO
RUN THERE.YOU CAN GET SOMETHING FROM THE
ENCLAVE,WHAT CODE ARE YOU
RUNNING.IT WILL PRODUCE A QUOTE SIGNED
BY THEPROCESSOR THAT SAYS THIS
IS THE CODE I’M RUNNING. FROM THE OUTSIDE YOU CAN VERIFY THAT THAT TRULY WAS AN INTEL GS PROCESSOR BY LOOKING AT THE
PUBLIC KEY FOR THAT PROCESSOR
AND THEN VERIFY IT IS THE CODE
YOU TRUST. AND ONCE IT IS,ESTABLISH A
SECURE CHANNEL WITH THAT CODE
AND HAND ITS KEYS SO THAT IT CAN
DECRYPT DATA INTO ITS ENCLAVE
AND THEN PROCESS ITSAFELY,PROTECTED BY THIS BLACK BOX.AND THE KINDS OF SCENARIOS — SO,ACTUALLY,FOR CONFIDENTIAL
COMPUTING, WE JUST ANNOUNCED THE
AVAILABILITY OF THE DC SERIES,SGX-ENABLED VIRTUAL MACHINES.YOU CAN LAUNCH THEM RIGHT NOW IN
EUROPE AND U.S.AND WE’RE ALSO ANNOUNCING THE
RELEASE OF AN SDK TO HELP YOU
WRITECONFIDENTIAL APPLICATIONS.ONE OF THE COOLEST SCENARIOS
THAT THIS OPENS UP IS SOMETHINGCALLEDMULTIPARTY MACHINE LEARNING.AND WHAT I’M GOING TO SHOW YOU
IS THEPOWEROF MULTIPARTY
MACHINE LEARNINGHERE.YOU KNOWWHAT . . .[ LAUGHING ]>>Mark: PROOF THAT THAT IS A
REAL DEMO.>>Mark: OKAY.SO, CONNECTING BACK TO OUR DEMO
VM.AND NOW WE’RE GOING TO GO INTOA
VIRTUAL MACHINE THAT IS ONE OF
THOSEDC SERIES VIRTUALMACHINES.AND THAT IS RIGHT HERE.SO,I’VE GOT TWO HOSPITALS.THEY’RE REPRESENTED BY DIFFERENT
DESKTOPS ANDTHEY BOTH HAVE
BREAST CANCER TRAINING DATA THEY
DON’T WANT TO SHARE WITH ONE
ANOTHER.BUT THEY’D LIKE TO COMBINE THEM
TO GET DEEPER INSIGHTS.THE FIRST THING THEY’RE GOING TODOISENCRYPT THEIR DATA TO THE
ENCLAVE.SO I’VE JUST ENCRYPTED IT.THEN THEY’RE GOING TOUPLOAD IT TO THE DC SERIES VM.AND WE’RE GOING TO DO THAT WITH
USER B AS WELL.HERE’S USER B’S DESKTOP.WE’RE GOINGTO GO TO DRAGTHIS,ENCRYPT IT,UPLOAD IT. WE’RE SIGNING IN AS USER B.AND THIS DATA IS ENCRYPTED SO
THAT ONLY THAT MACHINE LEARNING
ENCLAVE CAN SEE IT.NOW IT’S DOING TRAINING.WE’RE GOING TO SWITCH BACK TO
HOSPITAL A,BECAUSE THE GOAL
HERE IS TO GET BACKA TRAINED
MACHINE LEARNING MODEL. SO WE’RE GOING TO GO TO MODEL
AND WE’RE GOING TO DOWNLOAD IT.NOW WE’RE GOING TO GO TO
EVALUATE THE MODEL. AND WE SEE THAT THE LOCAL
MODEL — IF WE DO AN EVALUATION
ON THE LOCAL MODEL,WE HAVE AN
ACCURACY OF ABOUT 84.BUT THE MODEL WE JUST DOWNLOADED
FROM THE CLOUD FROM THAT COMBINED DATA SET MACHINE
LEARNING, THAT DATA FROM THE
OTHER HOSPITAL, YOU CAN SEEIT’S
97 ACCURACY.AT EITHER POINT,DID EITHER
HOSPITAL EXPOSE ITS DATA TO THE
OTHER ONE,OR TO ANYTHING ELSE,INCLUDING THE CLOUD? AZURE ADMINISTRATORS DON’T HAVE ACCESS,THE HYPERVISOR DOESN’T
HAVE ACCESS,THE HOSPITAL,OF
COURSE,DOESN’T HAVE ACCESS.AND SO THAT’S THEPROMISE OF
CONFIDENTIAL COMPUTING.[ APPLAUSE ]>>Mark: SO LET’S TALK ABOUT
STORAGE NOW.ANDAZURE STORAGE ARCHITECTURE IS TIERED ARCHITECTURE.SO YOU HAVE APIs ON THE FRONT
END.AND THOSE API SETS CONTINUE TO
GROW.SO BLOB WAS ONE OF THE FIRST
ONES THAT WE HAD.WE HAVE AZUREQUEUES ALSO BUILT ON TOP OF AZURE STORAGE,AZURE
FILES, AND THEN WE’VE GOT SMB ASWELL.AND THERE’SLOAD BALANCERS ON
THE FRONT END,LOAD BALANCING
TRAFFIC,SENDING THOSE REQUESTS
TO FRONT ENDS.THERE’S TABLE TIER OR PETITION
LAYER WHERE THE DATA ISPARTITIONED AND THEPARTITION MAPPING TO STREAM LAYER OR THE
EXTENT LAYER AT THE BOTTOM,THE
DISTRIBUTED FILE SERVICE ACROSS
THE SERVERS. SO A PARTICULAR BLOB MIGHT BE
BROKEN UP INTO CHUNKS THAT ARE
ON A BUNCH OF DIFFERENTSERVERS,AND EVERY CHUNK ISREPLICATED
THREE TIMES FOR HIGH AVAILABILITY.IF YOU USE A GRS STORAGE ACCOUNT,WHERE IT ALLOWS YOU TO
REPLICATE YOUR DATA INTO ANOTHER
REGION,THAT’S COPIED TO ANOTHER
STAMP ASYNCHRONOUSLY IN THAT
OTHER REGION.WE JUST ANNOUNCED RECENTLY AZURE
DATA LAKE STORAGE GEN TWO,A
MASSIVE SCALE-OUT DATA LAKE
SERVICE BUILT ON TOP OF AZURE STORAGE THAT INCLUDES ANHDFS
FRONT END,THE GEN TWO API.THIS IS SOMETHING THAT WE ADDED
ON TOP OF AZURE STORAGE. AND THE REASON THAT WE COULD
SCALE OUT THE WAY WE DID TO
SUPPORT THIS DATA LAKE TYPESCENARIO,WHICH IS MASSIVE
FILES,MASSIVE ACCOUNTS OF LOTS
OF BIG FILES,ISBYRE-ARCHITECTING STORAGE.TO TAKEWHAT USED TO BE
MONOLITHS OF A STORAGE STAMPWHERE THE FRONT THE END WAS
MAPPED TO THE PARTITION TABLE
ANDSTREAM SERVERS,TO SEPARATE THEM OUT.AND YOU CAN SEE THE KIND OF
SCALE THIS GIVES US FOR A BLOB, THE THROUGHPUT LIMITUP UNTILTHIS WAS 16 MEGABYTES PER
SECOND.WITH THISIT’S 50 GIGABYTES PER
SECOND. YOU CAN UPLOAD A FILE NOW THAT
USED TO TAKE 15 HOURS IN JUST A
FEWMINUTES. WE’VE ALSOEXPANDED OUR SSD
OFFERINGS. WE HAVE STANDARD HARD DISK,PREMIUM SSD,AND STANDARDSSD
WHICH HAS PERFORMANCE THAT’S
MORE LIKE A HARD DISK,BUT IT
HAS MUCH LOWERLATENCY.500IOPs, 60 MEGABYTES PERSECOND,SINGLE DIGIT LATENCY.AND THEN WE’VE ALSO INTRODUCED SOMETHINGCALLED ULTRA-SSD,WHICH WE JUST ANNOUNCED HERE AT
IGNITE. THIS IS OUR NEXT-GENERATION DISK STORAGE.THEPURPOSE-BUILT BLOCK STORAGE SERVICE ISCALLED DIRECT DRIVEBECAUSE THE HOSTS THAT THE DISKS
ARE MOUNTED TO HAVE INFORMATION
ABOUT WHICH SERVERS IN THE
DIRECT DRIVE CLUSTER HAVE THE
RELEVANT PIECES OF THEFILES. AND SO IT CAN GO TALKING
DIRECTLY TO THEM TO WRITE AND
READ DATA RATHER THAN GOING
THROUGH THE LOAD BALANCERS AND
THE FRONT END AND THE PARTITION TABLESERVERS THAT EVEN PREMIUM
SSD TRAFFIC HAS TO GO THROUGH.HOW MANY PEOPLE SAWCORY’S DEMO?A LOT OF YOU HAVEN’T SEEN IT.I’LL SHOW IT TO YOU REALLY
QUICKLY HERE.AND SO HERE I’VE GOTAN ULTRASSD SERVER,THESE SUPPORT UPTO
64 TERABYTE DRIVES. THIS HAMMERS THE DISK AS FAST AS IT CAN.NOW,THIS PARTICULAR VIRTUAL MACHINE IS PROVISIONEDFOR
160,000 IOPs,THE MOST YOU CAN
GET OUT OF ANY SINGLE DISK IN
ANY CLOUD.AND TO SHOW YOU HOW CONSTANT,HOW CONSISTENT THATPERFORMANCE,HERE’S THE IOPs PER SECOND.IT’S HOVERING RIGHT JUST AROUND160,000.AND THE LATENCY IS A MILLISECONDORLESS.THAT’SWHAT CORY SHOWED,WHICH
IS PRETTYIMPRESSIVE. BUT HOW ABOUT THIS?I’VE GOTA SPECIAL SERVER WITH A
SPECIAL DISK ON IT THAT WILL
SHOW YOU THAT THIS TECHNOLOGY IS
CAPABLE OF GOING HIGHER.SO WE’RE WORKING ON MAKING EVENBIGGER,FASTER DISCS.AND THIS IS A PREVIEW OF THAT THAT I WANTED TO SHOW YOU.I BEGGED THE TEAM SO I COULD
ONE-UPCORY.[ LAUGHING ]
>>Mark: SO LET’S FIRE THIS
ONE.>>Mark: THAT’S 250,000 IOPs.ANDTHE LATENCYISSTILLABOUT ONE MILLISECOND.[ APPLAUSE ]>>Mark: SO WE’RE GOINGEVEN
FURTHER WITHOUR DISKS THAN WE
HAVE,THAN WHAT I JUST SHOWED
YOU.OKAY.SO NOW I WANT TO TAKE YOUINSIDE
OF OUR DATA PLATFORMS AND TALK
ABOUT THE DATABASE ARCHITECTURES THAT WE’VE GOT.SO,WHEN YOU TAKE A LOOK ALL THE
WAY OUT,MANYOF OUR DATASERVICES,INCLUDING AZUREWAREHOUSE, HAVE THIS SIMILAR
TYPE OF HIGH-LEVEL ARCHITECTURE
WHERE YOU GO FROMREGIONS, EACH
REGION CONSISTS OF MULTIPLE DATACENTERS.THERE’S THE CONCEPT OFSTAMPS.THEY HAVE DEPLOYMENT GROUPS, A
NUMBEROFSERVERS,TYPICALLY ON
THE ORDER OF THOUSANDS.THOSE STAMPS HAVE FAULT DEMANDS,A SINGLE POINT OF FAILURE,GROUP
OF SERVERS WITH A COMMON SINGLE
POINT OF FAILURE. THAT WOULD BE A TOP OF THE RACK ROUTER OR A POWER DISTRIBUTION
UNIT. ONE OF THOSE FAILS,YOU LOSE
THOSE SERVERS.WE NEED TO MAKE SURE THE STORAGESERVICES UNDERSTAND FAULT DEMANDS TO SPREAD DATA OUT.I MENTIONED THREE COPIES OF
DATA.WITH THEPREMIUM AZURE DATABASE
DATABASE,YOU GETFOUR COPIES OF
DATA SPREAD OUT ACROSS DIFFERENT
FAULTDOMAINS.YOU CAN LOSE TWO SERVERS AND
STILL BE ABLE TO WRITE TOYOUR
AZURE DATABASE.WITHIN A MACHINE YOU CAN SEEAGENTS,CONTAINERS THAT ARERUNNING THOSE DATABASES, INSTANCES,THERE’S RESOURCE
GOVERNORS, THERE’S ATRANSPORTLAYER,THERE’S A MISSION CONTROL
THAT THROTTLES ON THEWAY IN,TRAFFIC SO THAT THE USER GETS
THE THROUGHPUT THAT THEY PROVISION.AND THEN THE DATABASE ENGINE
ITSELF.AND THE DATABASEENGINE, OF
COURSE,HAS A PROCESSOR,LANGUAGE RUNTIMES,AND SO ON.THAT’S THE OVERALL ARCHITECTURE.NOW,HOW MANYOF YOU HAVE USED
AZURE SQL DATABASE?LET’S SEE A RAISE OF HANDS.SO,QUITE A FEW OF YOU.HOW MANYOF YOU HAVE EVER WANTED
TO RAISE THE REQUEST UNITS ON
YOUR DATABASE ONLY TO EXPERIENCE THE FACT THAT YOU GOT DOWNTIME
AS THE DATABASE GETSRESIZED?ONE PERSON? I’M TALKING TO YOU,THEN.EVERYBODY ELSE CAN IGNORE THIS
SESSION.I’LL EXPLAIN THE KIND OF
EXPERIENCE THAT THE FEW PEOPLE
THAT RAISED THEIR HANDS HAVE.BECAUSE UP TO NOW,WHAT WE’VE
DONE IS CO-LOCATE THE DATABASE WITH THE DATA THAT IT MANAGES ON THE SAMESERVER.SO THERE’SMULTIPLE REPLICAS.IN THE CASE OF PREMIUM THERE’S
FOUR SERVERS,BUT ONE OF THEM IS
THE WRITE MASTER OUT OF THOSE
FOUR..IT WRITES TO ITS LOCAL DISK AND
SENDS THE DATA TO THE REPLICAS FOR THEM TO WRITE TO THEIR
DISKS. BUT YOU CAN’T SCALE UP THE
COMPUTE INDEPENDENT OF THE
STORAGE.THERE’S AONE-TO-ONE MAPPING
WITH REPLICAS AND DATA ON LOCAL
DISKS.SO WHAT WE’VE DONEIS
RE-ARCHITECTED AZURE DATABASE TO
BREAK THAT CONNECTION BETWEEN
COMPUTE AND STORAGE.THIS IS A TREND ACROSS DATA
SERVICES IN OUR CLOUD IN
GENERAL,TO SEPARATE COMPUTE ANDDATA,WHICH WILL ALLOW YOU TO
INDEPENDENTLY SCALE THE COMPUTE AND THE DATA.YOU CAN SHUT DOWN THE COMPUTE
AND STILL PRESERVE THE DATA, SOMETHING YOU COULDN’T DO WITH
THIS PREVIOUS ARCHITECTURE.HERE IS A DIAGRAM OF THIS NEW
ARCHITECTURE. WE CALLIT SOCRATES,BUT THE
OFFICIAL NAMEIS AZURE SQL
DATABASE FLEX SCALE. AND YOU CAN SEE THAT THERE’S THE COMPUTE TIER AT THE TOP.YOU CAN SEE THAT THERE’S FOUR
REPLICAS.ONE OF THEM IS THE WRITE MASTER.YOU CAN READ AND WRITE TO IT.THE OTHER ONES YOU CAN READFROM.IT SCALES IN READ.YOU CAN SCALE OUT THIS TIER. YOU CAN HAVE ANY NUMBER OF READ REPLICAS TO SCALE OUT THE AMOUNT OF DATA QUERIES THAT YOU CAN DO AGAINST THE DATABASE.NOW THE WAY THAT WE GET THIS
SEPARATION IS BY STORING THE LOG
ON A PREMIUM STORAGE SSD.SO THE LOG GETS SPIT OUT TO THESSD.THAT LOG IS READ FROM THESE STORAGE STAMPS THATDO THE
COMMITS OF THE TRANSACTIONS IN
THE DataStore. SOANY FAILURE AT ANY POINT,YOU
HAVE THE LOGS OR THE DATA INSTORAGE.AND NOW THE DATA IN STORAGE CAN
BE READ INTO ANY OF THOSE READ
OR WRITEREPLICAS FOR OPERATIONS
AND TRANSACTIONPROCESSING. SO LIKE I SAID,WE CAN TAKE THAT COMPUTE TIER AND SHUT IT DOWN
AND THAT DATA ISSTILL SITTING
IN STORAGE, OR TAKE THE COMPUTE
TIER AND SCALE IT OUT.AND THAT’S WHAT I’M GOING TO
SHOW YOU NOW.THIS IS THE FIRST TIME THIS HAS
BEEN SHOWN PUBLICLY. THE DYNAMIC SCALING OUT OF A
DATABASE. AND WHAT THIS ALSODOES IS LET
US BREAK THEDATABASE SIZE
BARRIER.SO THE MAXIMUM SIZE,ANYBODY
KNOW THE MAXIMUM SIZE OFAN
AZURE SQL DATABASE TODAY,ANYBODY?YOU CAN ASK CORTANA AND SHE’LL
PROBABLY TELLYOU.FOUR TERABYTES.SO WHAT I’M GOING TO SHOW YOU
HERE,HERE ISMY SOCRATES
DATABASE. ANDIT’S 50 TERABYTES.AND WE’RE GOING TO A HUNDRED.SO THIS IS WHERE WE ARE RIGHT
NOW.WE’RE WORKING ON SCALING. SO WE’REAT 50,WANTED TO GIVE
YOU A PREVIEW.THIS IS A PREMIUM L1 DATABASE WITH 250DTUs.THOSE ARE THE AMOUNT OF QUERY
PROCESSING THAT THIS DATABASE WILL SUPPORT.IN THE PAST,LIKE I MENTIONED,IF YOU HAD AN L1 TO CREATE AN
L2,WHAT HAD TO HAPPEN WAS YOU HAD TO SHUT DOWN THAT L1,CREATE THE L2 ONA BIGGER SERVER AND
THEN COPY THE DATA OFF OF THE L1
ON TO THEL2,AN OPERATION WHICH
COULD TAKE HOURS.INA50-TER BYTE DATABASE,THAT’S GOING TO BE A WHILE.WE WANTED TO SCALE WHILE THIS
DATABASE WAS RUNNING AND WITHOUT
HAVING TO COPY THE DATABASE. SO WHAT I’M GOING TO DOIS GOINTO . . .THISIS . . .NOT
THERE.I’M GETTING LOST IN MYDEMOS.SO THAT IS OVER HERE.HERE WE GO.SO,THIS IS THE — I’M GOING TO
SENDA SQL COMMAND TO SCALE THAT
UP TO AN L5, EXECUTE THAT.OOPS.SELECT THEWHOLE LINE.AND THIS IS GOING TO TAKE ABOUT TWO MINUTES.SO WE’RE GOING TO COME BACK AND
LOOK AT THAT IN JUST A COUPLE
MINUTES.BUT WE’RE GOING TO SEE THIS
THING DYNAMICALLY SCALE OUT
WITHOUT THE HOURS THAT IT WOULD HAVE TAKEN OTHERWISE.SO,FINALLY, I’M GOING TO
CONCLUDE BY TAKING A LOOK AT
COSMOSDB.COSMOS D IS A RELATIVELY NEW NO
SQL DATABASE.I THINK IT REALLY REPRESENTS THE
PINNACLE OF — KIND OF THE
CUTTING EDGE OF CLOUD-NATIVEDATABASES.IT’S GOT A NUMBER OF DIFFERENT
CHARACTERISTICS WHICH ARE UNIQUE
TO IT IN THE PUBLIC CLOUD.FOR ONE,IT’SA MULTIMODAL
DATABASE THAT SUPPORTS A NUMBER
OF DIFFERENT APIs FROM THE
AZURE TABLES API TO A SQLJSONTO MONGO DB.YOU’RE LIKELY TO SEE EVEN MORE.BESIDES SUPPORTING MULTIPLE
APIs,IT ALSO SUPPORTS
MULTIPLE CAN I CAN I CONSISTENCY LEVELS.ON A STANDARD DATABASE YOUGET
STRONG CONSISTENCYBUT SACRIFICE
PERFORMANCE. IT WAS SAYING WE’RE GOING TO BE INCONSISTENT AND THAT WILL LET
US SCALEBETTER.COSMOS DB LETS YOU GO FULLY
ASYNCHRONOUS AND HAS THREE LEVELS IN BETWEEN,INCLUDING THE DEFAULT,SESSION CONSISTENCY,TYPICALLY WHAT MOST APPLICATIONSNEED,WHICH IS ANYBODY
PARTICIPATING IN THIS SESSION
WITH A DATABASE SEES THE READS
AND THE WRITES IN THE SAME ORDER.NOW,DIFFERENT APPLICATIONS HAVE
DIFFERENT SESSIONS THAT AREN’T INTERACTING WITH ONE ANOTHER.THEY CAN BE OPERATING IN
PARALLEL AND THAT ALLOWS YOUR
DATABASE TO GET MUCH HIGHER
PERFORMANCE,EVEN THOUGH IT’S
TECHNICALLY INCONSISTENT.IT’S ALSO DESIGNED FOR A MASSIVE
SCALE. IT’S DESIGNED TO WORK ACROSSREGIONS. IT’S THE ONLY DATABASE IN THE
PUBLIC CLOUD THAT SUPPORTS SLAS
ACROSS FIVEDIMENSIONS —
LEGACY,CONSISTENCY, THROUGHPUT,AVAILABILITY,AND THERE’S A
FIFTH ONE.[ LAUGHING ]>>Mark: IT’S A HOMEWORK
ASSIGNMENT.BUT IF YOU TAKE A LOOK AT THE
COSMOS DB ARCHITECTURE,IT
SPLITS DATABASES INTO PARTITIONS
THAT ARE SPLIT INTO MULTIPLE —
REPLICATED IN MULTIPLE REGIONS.AND SO YOU CAN REPLICATE LITERALLY TO EVERY REGION IN
AZURE A SINGLEDATABASE.AND WITHIN A REGION THAT’S GOT
REPLICATION TURNED ON, YOU CAN
SEE THAT THERE’S A VIRTUAL REPLICA.WE SAW IN THE SQL DATABASE THERE
WAS FOUR REPLICAS READING FROM
LOCAL.IN THE CASE OF COSMOS DB,THERE’S A SINGLE WRITE MASTER BUT THERE’S ALSO A PROXY TO
ANOTHER REGION THAT PARTICIPANTSPARTICIPATES IN WRITING DATA TO OTHER
REGIONS. THAT IS THE WAY THAT COSMOS DB
WORKS ACROSSREGIONS.TO SHOW YOU JUST HOW INCREDIBLY
SCALABLE COSMOS DB IS, I’M GOING
TO SHOW YOUA DEMO THAT SHOULD
DRIVE THE POINT HOME PRETTY WELL.LET ME SEE HOW MY FLEX DB IS GOING.THERE WE GO,PREMIUM L5.WITH 4,000DTUs.WE SCALED OUTTHAT DATABASE FROM
250 TO 5,000DTUs,SOMETHING WITHA 50 TERABYTE DATABASE
WOULD HAVE TAKEN HOURS,IN JUST
A FEW MINUTES. SO THAT’S COSMOSDB.[ APPLAUSE ]>>Mark: NOW,I’VE GOT ONE
LAST DEMO TO SHOW YOU,AND
THAT’S THIS COSMOS DB DEMO.>>Mark: AND HERE’S THE PORTAL.SO WHAT I’VE GOT RIGHT HERE IS
AN EXAMPLE OF AN APP.IT SELLS TICKETS.IT’S A TICKET-SELLING APP. AND WE’VE GOT ONE REGION THAT IS ACTIVE IN WRITING TO THE
DATABASE RIGHT NOW.SO WE’VE GOT THIS GUI IS
BASICALLY A LOADGENERATOR
SIMULATING PEOPLE BUYING
TICKETS.WE ONLY NEED ONE REGION.WE’RESELLING 11,000 TICKETS PER
MINUTE.YOU CAN THINK OF THAT AS
TRANSACTIONS PER MINUTE.THIS DATABASE ISA 100 TERABYTESCALE,SHOWS THE SIZE DATABASES
CAN GET TO.YOU CAN SEE THE LATENCY
REPORTING ACROSS ALL THEREGIONS,THE AVERAGEIS SIX MILLISECONDS,TO NINE.UNDERTHE TEN MILLISECONDS SLA
THAT COSMOS DB HAS FOR READS AND
WRITES.IT WILL ALWAYSBE,AS OURPROMISE,UNDER TEN MILLISECONDS.BUT LET’S SAY THAT WE’VE GOT A
SURGE IN TICKETSALES. SO I’M GOING TO GO INTO PANICMODE, WHICH IS SCALING UP THE
DATABASE RIGHTS.WE’RE SCALING TO A TOTAL OF
THREE REGIONS.IT LOOKS LIKE FOUR BECAUSE THIS
IS A WRAP AROUND SO THAT YOU CAN
SEE THAT THIS ASIAN ASIA IS THE SAME
AS THIS ASIA.YOU CAN SEE THE LATENCY OF
RIGHTS IN THATREGION,EIGHTMILLISECONDS,FIVE. SO WE’VE GOT RIGHTS COMING FROM ALL OVER THE WORLD,EFFECTIVELY
TICKET SALES.YOU CAN SEE OUR LATENCYAVERAGE. WE’RE DRIVINGNOW 500,000TRANSACTIONS PER MINUTE.AS TICKET SALES UPDATING THIS
DATABASE.HOW’S THAT?[ APPLAUSE ]
>>Mark: OKAY.I’M GLADYOU FOUND THAT,KIND
OF,IMPRESSIVE.SO I’M GOING TO TAKE IT UP A
NOTCH.WE’RE GOING TO GO INTO EXTREME
PANICMODE,BECAUSE SCOTT
GUTHRIE IS GOING TO BE AT THIS
EVENT AND NOW TICKET SALES ARE
GOING CRAZY.WE’RE GOING TO SCALE OUT NOW TO
20REGIONS.AND WE’RE GOING TO SUPPORT,AS
YOU CAN SEE THIS SCALING UPHERE . . .782,000 TICKET PERMINUTE, 1 MILLION TICKETS PER
MINUTE ACROSS — GOING FROM 20REGIONS,MULTI-MASTER WRITE
CAPABILITY,SO THE WRITES CAN HAPPEN CLOSE TO WHERE THEY ARE
BEING CREATEDATBELOW TENMILLISECONDS.TO SHOW YOU HOW YOU’D CONFIGURE
THAT INTHE PORTAL,THAT WAS MYCONNECTION ERROR FROM
DISCONNECTING TO THENETWORK. THIS ISTHE ACTUALCOSMOS —
WHAT? [ LAUGHING ][ CLAPPING ]>>Mark: THAT’S NOT IT.I THINK THIS IS IT. HOPEFULLY I HAVE ACCESS.NO,I DON’T THINK THIS IS THE
ONE.I THINK IT’S INA
DIFFERENT . . .IFI GO TO
REPLICATE DATA LOCALLY,GLOBALLY,WE SHOULD SEE 20
REGIONS.THIS IS NOT IT.I DON’T THINK IT’S THIS ONE,EITHER. SO MAYBE I’M LOGGED OUT NOW.NOPE, THAT’S NOT IT,EITHER.BUT WHAT YOU WOULD DO IS SELECT
THE REGIONS THAT YOU WANT TO
REPLICATE TO.IT’S AS SIMPLEAS SELECTING
THEM.IF I WENT TO THE OVERVIEW ON THE
DATABASE THAT I WAS JUST
HITTING,THE NUMBER OF
TRANSACTIONS THAT THATDATABASE HAS PROCESSED,BECAUSE WE
STARTED MESSING WITH THIS
YESTERDAY SO THAT’S BEEN ONLINE
SINCE YESTERDAY,WE’VE BEEN
RUNNING THIS TEST UP AND DOWN, IS 1.5 TRILLION TRANSACTIONS
SINCE YESTERDAY,IS WHAT YOU
WOULD HAVE SEEN ACROSS 20REGIONS,TEN MILLISECOND LATENCY
ORLOWER,MULTI-MASTER WRITE CAPABILITY.THAT’S WHY COSMOS DB IS AT THE
CUTTING EDGE OF WHERE DATA
SERVICES ARE IN THE CLOUD.WITH THAT,I WANT TO WRAP IT UP.I HOPE THATYOUFOUND THISINTERESTING,AND HOPE IT GAVE
YOU A LITTLE LOOK BEHIND THE
COVERS OF WHAT OUR
INFRASTRUCTURE LOOKS LIKE,AS
WELL AS SHOWED YOU SOME OF THE
COOL INNOVATIONS,MANY OF WHICH
WE HAVE COMING OUT HERE, AND
THEN I GAVE YOU A PREVIEW OF DIRECTIONS WE’RE GOING.WITH THAT, I WANT TO THANK YOU
VERY MUCH.HAVE

Leave a Reply

Your email address will not be published. Required fields are marked *