DB118 - E2E machine learning for enterprises in a secure way
Las características de seguridad y gobierno como el aislamiento de la red (VNET, IP privada), la identidad administrada, RBAC, Policy, etc. son lo más importante para casi todas las empresas del mundo. Pero crear una solución ML de manera segura y que cumpla con los requisitos de cumplimiento empresarial es un desafío. Con Azure Machine Learning, vea cómo crear y entrenar modelos de aprendizaje automático con estaciones de trabajo alojadas en la nube administradas y seguras (instancias de cómputo). Finalmente, implemente modelos en clústeres privados de AKS para completar el ciclo de vida de ML.
Transcript
00:03>> SO WELCOME. I'M JUMPEI HIRONO,
00:07PROGRAM MANAGER, AND WE ALSO HAVE
00:12CO-PRESENTER VIJETA JOHRI, ANOTHER
00:13PROGRAM MANAGER ON THE PLATFORM
00:14TEAM. SO I HAVE BEEN WORKING FOR
00:18MACHINE LEARNING AND DEEP LEARNING
00:19INDUSTRY MORE THAN FOUR YEARS, AND
00:22I SHOW MANY ENTERPRISE CUSTOMERS
00:24STARTED USING OR STARTED BUILDING
00:27MACHINE LEARNING PLATFORM THAT BUILDING
00:32A MACHINE LEARNING PLATFORM IN A
00:33COMPLIANT WAY IS SUPER CHALLENGING
00:36BECAUSE BASICALLY COMPANY NEEDS
00:38TO CRAFT LOTS OF OPEN SOURCE TECHNOLOGY
00:41AND MULTIPLE COMPUTING WITNESSES.
00:47DATA SCIENTISTS ARE ALSO IN THAT
00:48CHALLENGE. SO HAVING SAID THAT,
00:51IN THIS SESSION WE COVER HOW TO
00:54BUILD SECURE AND COMPLIANT MACHINE
00:56LEARNING PLATFORM USING AZURE CLOUD
00:58AND AZURE MACHINE LEARNING PLATFORM.
01:00SO HAVING SAID THAT, YEAH, LET'S
01:03GET STARTED WITH THAT PRESENTATION.
01:05SO AT FIRST, YEAH, LET ME COVER
01:08THE AZURE MACHINE LEARNING FUNDAMENTALS
01:09AND WHAT KIND OF RESOURCES AZURE
01:12MACHINE LEARNING USED, AND AFTER
01:15THAT I WILL COVER KEY CONSIDERATIONS
01:17SUCH AS IDENTITY, DATA AND NETWORK.
01:20AND THEN VIJETA WILL DO A DEEP DIVE
01:22ON COMPUTE INSTANCE, AND AFTER THAT
01:24WE WILL DO END-TO-END SECURE AND
01:28COMPLIANT MACHINE LEARNING PLATFORM
01:29DEMO. LASTLY WE SHARE SOME INSIGHTS
01:32ON ROAD MAP AND Q&A. SO LET'S GET
01:37STARTED. SO BASICALLY AZURE MACHINE
01:41LEARNING IS SERVICE TO BUILD MACHINE
01:42LEARNING PLATFORM ON TOP OF AZURE.
01:45SO WE HAVE MANY DEPENDENT RESOURCES
01:46ON AZURE. SO LET ME EXPLAIN ONE
01:49BY ONE. SO FIRST ONE IS ASSOCIATED
01:53WITH -- SO IF YOU AZURE MACHINE
01:56LEARNING WORKSPACE YOU WILL GET
01:58THESE RESOURCES AND STORAGE FOR
01:59STORING YOUR DATA, KEYBOARD STORING
02:02YOUR SECRETS -- STORING DOCKER IMAGE.
02:10SO FUNDAMENTAL RESOURCES TO DESIRE
02:12MACHINE LEARNING. SO AFTER THAT
02:15YOU NEED TO TRAIN AI MODEL OR AZURE
02:16MACHINE LEARNING PLATFORM. SO WE
02:18HAVE TWO TYPES OF RESOURCES. FIRST
02:22ONE IS COMPUTE INSTANCE FOR THE
02:26CLOUD-BASED WORKSTATION FOR DATA
02:29SCIENTIST SO FOR MODERN DEVELOPMENT.
02:32AND AFTER MODEL DEVELOPMENT DATA
02:35SCIENTISTS WILL SUBMIT A TRAINING
02:37JOB WHICH IS A SCALEABLE TRAINING
02:40ENVIRONMENT. AND AFTER MODEL TRAINING,
02:42YOU WILL DEPLOY MODEL TO CLOUD.
02:46SO IT IS CALLED INFERENCE IN DEEP
02:49LEARNING WORLD, SQUALLING IN MACHINE
02:52LEARNING WORLD, SO WE SUPPORT THEM
02:54FOR THAT. AND ADDITIONALLY YOU MIGHT
02:58WANT TO MOUNT ADDITIONAL STORE SUCH
03:01AS SQL SERVER. THESE FOUR ARE THE
03:04REQUIRED RESOURCES FOR AZURE MACHINE
03:05LEARNING PLATFORM. SO IF YOU THINK
03:07ABOUT SECURE AND COMPLIANT DESIRE
03:10AZURE MACHINE LEARNING, WE NEED
03:11TO THINK ABOUT THESE AS WELL. SO
03:12HAVING SAID THAT, LET ME MOVE ON
03:16TO THE SECURITY CONSIDERATIONS,
03:17IDENTITY, DATA AND NETWORK. AND
03:20FIRST ONE IS IDENTITY. SO IDENTITY
03:23HAS TWO SIDES, AUTHENTICATION AND
03:26AUTHORIZATION. SO FOR THE AUTHENTICATION,
03:28YEAH, WE HAVE BIG COVERAGE OF AUTHENTICATION
03:31MECHANISM. SO FOR WORKSPACE CREATION,
03:35WORKSPACE ACCESS, WE SUPPORT MULTIPLE
03:38AUTHENTICATION AZURE CLI AUTHENTICATION
03:43-- IDENTITY AND FOR THE TRAINING,
03:46YEAH, BASICALLY DATA SCIENTIST WANT
03:48SSI ACCESS TO COMPUTER -- SO WE
03:50SUPPORT THAT WITH STORING SECRETS
03:53IN KEYBOARD. AND FOR SCORING AKS
04:02TOKEN AUTHENTICATION IS SUPPORTED.
04:05AND LASTLY COMPUTE INSTANCE -- SO
04:08THAT'S THE AUTHENTICATION, AND NEXT
04:12ONE IS AUTHORIZATION, IN OTHER WORDS,
04:16RBAC. AZURE MACHINE LEARNING PROVIDES
04:19GRANULAR RBAC FOR ROLES SUCH AS
04:21DATA SCIENTISTS AND DATA ENGINEERS.
04:26SO THE EXAMPLE OF DATA SCIENTIST,
04:30SO LET'S SAY -- CAN DO EVERYTHING
04:32BUT WE DO NOT WANT DATA SCIENTIST
04:37TO CREATE ADDITIONAL -- SO IF YOU
04:39LOOK AT THE CREATION ON THE LEFT-HAND
04:41SIDE, CAN YOU SEE DELETE BUTTON,
04:44BUT IF YOU LOOK AT THE SCREEN SHOT
04:47ON THE RIGHT-HAND SIDE, IT IS A
04:48DATA SCIENTIST'S VIEW. YOU CANNOT
04:52SEE THE NEW DELETE BUTTON. SO IT'S
04:54A PART OF GRANULAR RBAC, AND BY
04:58CREATING CUSTOM, THOUGH, FOR OTHER
05:01DATA SCIENTISTS, WE CAN APPLY THIS
05:03KIND OF RBAC. AND WE WILL CREATE
05:07BUILT INDOOR FOR TYPICAL ROLES IN
05:10THE NEAR FUTURE. SO THAT'S THE IDENTITY.
05:14AND FOR THE DATA PROTECTION, YEAH,
05:16AZURE MACHINE LEARNING SUPPORTS
05:18ENCRYPTION IN TRANSIT AND AT REST,
05:20AND USING BOTH MICROSOFT MANAGED
05:22KEY, WHICH IS DEFAULT, AND YOU CAN
05:26ALSO CONFIGURE CUSTOM MANAGE YOUR
05:28KEY. IN THE CASE OF CUSTOM MANAGED
05:30KEY, ADDITIONAL RESOURCE GROUP IS
05:33CREATED FOR CUSTOMER TO STORE METADATA
05:38AND MACHINE LEARNING. LAST ONE IS
05:41THE TOOL SECURITY. I KNOW THAT YOU
05:43THINK ABOUT THAT AND WE NEED TO
05:46UNDERSTAND THE NETWORK COMMUNICATION
05:48BETWEEN RESOURCES. SO WE HAVE THREE
05:50BIG RESOURCES, SO AZURE MACHINE
05:52LEARNING WORKSPACE, IT IS MANAGED
05:55BY MICROSOFT. AND YOU ALSO HAVE
05:59-- INSIDE YOUR SUBSCRI SUBSCRIPTION,
06:02AND YOU ALSO HAVE ASSOCIATED RESOURCES
06:07SUCH AS -- IT IS ALSO INSIDE YOUR
06:13SUBSCRIPTION. AND SO WE HAVE THREE
06:16COMMUNICATION -- SO LET ME EXPLAIN
06:18ONE BY ONE. SO FIRST COMMUNICATION
06:20IS HERE, SO YOU ARE COMPUTE RESOURCES
06:24NEEDS ACCESS TO AZURE MACHINE LEARNING
06:26WORKSPACE, AND IT HAPPENS OVER PUBLIC
06:29IP BY DEFAULT. BUT YOU CAN CHOOSE
06:34ENABLED WORKSPACE, SO WE START COMMUNICATION
06:38OVER PRIVATE IP. AND IN IGNITE WE
06:42ANNOUNCED GA THE WORKSPACE -- SO
06:46IT WILL BE AVAILABLE IN ALL PUBLIC
06:51AZURE MACHINE LEARNING REGIONS EXCLUDING
06:54CHINA. SO PLEASE TRY THAT. SO FOR
06:59YOUR COMPUTE RESOURCES NEED ACCESS
07:01TO ASSOCIATED RESOURCES. AND NOW
07:03WE SUPPORT SERVICE END POINT AND
07:06WE WILL SUPPORT PRIVATE END POINT
07:09SOON. SO IF YOU WANT TO LIMIT ACCESS
07:13ONLY FROM YOUR VNET SERVICE END
07:15POINT, PLEASE CONTINUE YOUR SERVICE
07:18END POINT. THAT SERVICE END POINT
07:21IS SERVED AS PUBLIC ID. IF YOU PREFER
07:25PRIVATE, THEN PLEASE CONFIGURE PRIVATE
07:27END POINT FOR THAT. WE ALSO HAVE
07:29REMAINING COMMUNICATION PASS, BUT
07:31THIS GOES THROUGH A MICROSOFT BACKBONE
07:34OVER PRIVATE IP, AND WE ALSO NEED
07:38TO THINK ABOUT COMMUNICATION INSIDE
07:40THE NET, SUCH AS COMPUTING AND COMPUTE
07:46CLUSTER, AND THIS NETWORK ACTUALLY
07:48SHOWS, BEHIND NET IS SUPPORTED,
07:51BUT EVEN WE PUT COMPUTE RESOURCES
07:54BEHIND VNET, THEY HAVE PUBLIC IP.
07:59AND SO SOME -- SO IT WILL BE EXPLAINED
08:02BY VIJETA THAT IT IS -- AND FOR
08:09THE MODEL -- SO HAVING SAID THAT,
08:14YEAH, I COVERED KEY SECURITY CONFIGURATIONS
08:18SUCH AS IN IDENTITY, DATA AND NETWORK.
08:22SO THEN LET'S DO THE DEEP DIVE ON
08:24COMPUTE STANCE BY VIJETA. SO VIJETA,
08:27CAN YOU PLEASE? >> SURE. THANK YOU,
08:29JUMPEI. HELLO, EVERYONE. SO WE THINK
08:34OF THE INSTANCE AS THE MANAGED WORKPLACE
08:37WORKSTATION FOR DATA SCIENTISTS
08:38WHICH MAKE IT EASY FOR THEM TO GET
08:41STARTED WITH MACHINE LEARNING DEVELOPMENT
08:43BY PROVIDING ADMIN MANAGEMENT AND
08:45READINESS CAPABILITIES. ALL THE
08:47RESOURCES ARE CREATED IN -- SO THAT
08:49IS WHY WE THINK WE OFFER IT AS A
08:51PAST SERVICE. AND IT'S REALLY AN
08:55ATTRACTIVE DEVOPS ENVIRONMENT. THEY
08:56CAN RUN NOTEBOOKS THROUGH THEIR
08:59TOOLS SUCH AS JUPYTER LAB, JUPYTER,
09:03AZURE MACHINE LEARNING NOTEBOOKS,
09:04WHAT WE HAVE IN THE LEFT PANE, AND
09:06THEY CAN RUN SCRIPTS THROUGH R STUDIO.
09:11IT'S WINDOWS 16 AND THEY ARE LOOKING
09:14TO UPGRADE IT. IT IS FULLY CUSTOMIZABLE
09:18DATA CENTERS. THEY CAN INSTALL ANY
09:21PACKAGES OR DRIVERS THEY WOULD NEED
09:23FOR THEIR WORK. SO THEY HAVE FULL
09:27FLEXIBILITY THERE. -- MACHINE LEARNING
09:30WORKSPACE AND PROVISIONED AS A CHILD
09:34RESOURCE OF MACHINE LEARNING WORKSPACE.
09:38IT IS ALSO A TRAINING TARGET WHERE
09:42USERS CAN SUBMIT LOCAL TRAINING
09:44JOBS INCLUDING GPU DISTRIBUTOR TRAINING
09:48JOBS, WHICH WE WILL DEMO LATER ON,
09:51AND WE SAW JUST A COUPLE OF MINUTES
09:55AGO WE CONTINUE TO INVEST ON RELIABILITY
10:00AND PERFORMANCE. SO IN TERMS OF
10:04BY THE END OF SEPTEMBER IT WILL
10:04ENTERPRISE READINESS CAPABILITIES,
10:07WE HAVE VIRTUAL NETWORK SUPPORT
10:08WHICH IS REALLY NEEDED BY EVERY
10:16ENTERPRISE FOR NETWORK. AS WE WILL
10:17SHOW IN THE DEMO, AND YOU CAN USE
10:21ARM TEMPLATES, REST TO HELP AUTOMATE
10:24CREATION AND MANAGEMENT. WE ALSO
10:27SUPPORT ADMIN CREATE ON BEHALF OF
10:29DATA SCIENTISTS AND ASSIGN TO THEM
10:30FOR USE. WITH THIS THERE'S A CLEAR
10:33SEPARATION OF CONCERNS WHERE THE
10:35CENTERS DON'T NEED TO HAVE COMPUTE
10:37PERMISSIONS. AND WE SUPPORT RBAC
10:39AND THROUGH RBAC YOU CAN CONTROL,
10:41CREATE, START, STOP, RE-START, DELETION,
10:44AND WHO CAN ACCESS THE NOTEBOOKS.
10:47WE SUPPORT ENCRYPTION IN TRANSIT,
10:51AND THERE IS A CONFIGUREABLE ACCESS
10:53USING PUBLIC-PRIVATE KEY MECHANISM.
10:59IN TERMS OF DATA SCIENTISTS OR PRODUCTIVITY
11:02CAPABILITIES, EACH INSTANCE IS LIKE
11:05A PRE-CONFIGURED MACHINE LEARNING
11:08ENVIRONMENT WITH THE MOST POPULAR
11:10PACKAGES, DRIVERS AND REALLY ANYTHING
11:12THAT DATA SCIENTISTS NEED TO GET
11:15STARTED AND BE ALREADY HAVE CREATED
11:17THE CONTENT ENVIRONMENTS FOR THE
11:19SCIENTISTS TO RE-USE. YOU CAN USE
11:23ANY OF THE VM FAMILIES AND CAN COLLABORATE
11:25ON THE NOTEBOOK FILES IN THE SAME
11:28WORKSPACE. AS I MENTIONED, EACH
11:35-- SHARE STORAGE, WORKSPACE STORAGE.
11:40WE RECENTLY RELEASED AZM EXTENSION
11:42THROUGH WHICH YOU CAN RUN YOUR GENERAL
11:45BOOKS USING GOVERNANCE AS A REMOTE
11:48SERVER, AND THIS CAPABILITY DOESN'T
11:50NEED ANY SSH ENABLING. IT HAS A
12:01JOB SCHEDULER, IT IS BACKED BY JOB
12:04QUEUE AND ALSO IS ABLE TO RUN MULTIPLE
12:06JOBS IN PARALLEL. FOR TRAINING WE
12:11SUPPORT POPULAR FRAMEWORKS. WE ALSO
12:14HAVE -- WHEN USING AZURE MACHINE
12:16LEARNING WITH NOTEBOOKS, REDUCING
12:19THE NEED FOR AUTHENTICATION AND
12:22ENSURING CONTINUITY WE SUPPORT NOTEBOOK
12:24SESSIONS BY PREVENTING SESSION TIMEOUTS.
12:28>> OKAY, SO YEAH, LET'S GET INTO
12:31DEMO. SO AT FIRST LET ME EXPLAIN
12:35THE DEMO ENVIRONMENT. SO MACHINE
12:41LEARNING WORKSPACE IN HERE. I CREATE
12:45END POINT WITHIN YOUR FREE NET,
12:47AND I ALSO CREATE PRIVATE END POINT
12:51FOR ASSOCIATED -- SUCH AS STORAGE,
12:55AND WE HAVE BUNCH OF TRAINING AND
12:58INFERENCE RESOURCES, SUCH AS COMPUTE
13:00INSTANCE AND AKS. AND YOU KNOW THAT
13:04YOU EMULATE YOUR ACCESS FROM OUR
13:08OWN PREMISE ENVIRONMENT, WE CREATED
13:10A GATEWAY AND WE ACCESSING TO THIS
13:14DEMO ENVIRONMENT FROM OUR PRESENTATION
13:16PC. OF COURSE EVERYTHING IS ENCRYPTED
13:20AND RBAC -- IS APPLIED. SO LET'S
13:23MOVE ON TO THE DEMO ENVIRONMENT.
13:25SO THIS IS THE RESOURCE GROUP WHICH
13:28HAS PRIVATELY ENABLED WORKSPACE.
13:32SO THIS IS THE WORKSPACE. WHOOPS.
13:36I AM REFRESHING, YEAH, SOMETIMES
13:41HAPPENS. SO YES, THIS IS AZURE MACHINE
13:45LEARNING WORKSPACE. SO LET ME LAUNCH
13:48A MACHINE LEARNING PORTAL. PLEASE
13:51NOTE THAT I HAVE NOT ESTABLISHED
13:52A VPN CONNECTION SO I CANNOT ACCESS
13:55STUDIO WORKSPACE. SO LET ME CLOSE
14:00THIS FOR NOW AND LET ME COLLECT
14:03THE -- LET ME ESTABLISH THE VPN
14:07CONNECTION TO THE VNET WHICH HAS
14:09AZURE MACHINE LEARNING WORKSPACE.
14:11SO IT IS CONNECTING AND CONNECTED.
14:14SO GO BACK TO THE BROWSER AND LET
14:18ME LAUNCH MS STUDIO AGAIN. SO NOW
14:22I CAN ACCESS STUDIO, MACHINE LEARNING
14:25WORKSPACE AND I CAN SEE A BUNCH
14:28OF EXPERIMENT AND COMPUTE CLASSES.
14:31AND YEAH, GO BACK TO THE AZURE PORTAL
14:35AND LET ME CHECK THE CONFIGURATION
14:38OF MACHINE LEARNING WORKSPACE AND
14:40ASSOCIATED RESOURCES. SO AT FIRST,
14:41YOU KNOW, AS I SAID, I CREATED THE
14:45PRIVATE END POINT FOR AZURE MACHINE
14:48LEARNING WORKSPACE. SO IF I CLICK
14:50HERE, SO YOU CAN SEE THE PRIVATE
14:53IP ADDRESSES. I USE THIS PRIVATE
14:58IP ADDRESS TO USE STOODIO AZURE
14:59MACHINE LEARNING WORKSPACE. AND
15:02IF I GO BACK TO THE RESOURCE CLIP
15:03AND, YEAH, AND LET ME CHECK THE
15:06COMPUTATION OF ASSOCIATED RESOURCES.
15:09FOR EXAMPLE, STORAGE. SO IF I CLICK
15:11STORAGE AND GO TO THE PRIVATE END
15:13POINT SECTION, YEAH, AS YOU CAN
15:15SEE, I CREATED THE TWO PRIVATE END
15:17POINTS. ONE IS FOR -- AND ONE --
15:22SO WE -- EVERY ACCESS TO STORAGE
15:25IS OVER PRIVATE IP. IF I GO TO THIS
15:29SECTION, YOU CAN SEE I HAVE NOT
15:31CREATED SERVICE END POINT IN HERE.
15:34AND GO BACK AGAIN TO MACHINE LEARNING
15:37PORTAL AND GO TO THE COMPUTE SECTION.
15:38YOU CAN SEE I CREATED OR VIJETA
15:42CREATED COMPUTE INSTANCES, AND WE
15:46HAVE A BUNCH OF COMPUTER [INDISCERNIBLE]
15:48NO PUBLIC ID CLUSTERS AND WE ALSO
15:51HAVE AKS CLUSTER, RIGHT? SO GO BACK
15:55TO MY PRESENTATION, SO AS CONFIRMED
15:58EVERYTHING SET UP PRIVATE END POINT,
15:59PRIVATE END POINT AND YOU CAN SEE
16:01THE RESOURCES AND CONNECTIONS THROUGH
16:03THE VPN GATEWAY. THEN MOVING ON
16:07TO THE SECOND PHASE, DEEP DIVE COMPUTE
16:09INSTANCE, VIJETA, CAN YOU START
16:15THE DEMO? >> SURE. THANKS, JUMPEI.
16:19SO I AM ALSO IN THE SAME VNET WHICH
16:23JUMPEI SET UP, AND THEN [INDISCERNIBLE]
16:25DEPLOYED, AND MY LAPTOP IS CONNECTED
16:28TO THE SAME VNET THROUGH THE VPN
16:32CONNECTION, SO THAT'S WHY I'M ABLE
16:33TO ACCESS THE WORKSPACE, SAME AS
16:35JUMPEI. WHEN YOU COME TO THE COMPUTE
16:38PANE ON THE LEFT SIDE, THIS IS WHERE
16:40YOU SEE THE -- ALL THE -- COMPUTES
16:43AND INSTANCES AND -- ARE THE MAPPED
16:45COMPUTES, AND -- MANAGED COMPUTES,
16:47AND HERE IS WHERE THEY ARE ALL LISTED
16:50ALONG WITH THE URIS FOR EACH OF
16:52THE LABS WE TALKED ABOUT, JUPYTER
16:57LAB AND STUDIO. ALSO WANT TO TAKE
16:58A BRIEF MOMENT TO JUST LOOK AT A
17:00COMPUTE CLUSTER. SO THIS IS A COMPUTE
17:03CLUSTER WHICH IS DEPLOYED IN THE
17:06PRIVATE WORKSPACE, AND WHAT IS INTERESTING
17:10ABOUT THIS CLUSTER IS DOESN'T USE
17:11ANY PUBLIC IP FOR COMMUNICATION,
17:13WHICH IF YOU SEE HERE, THE PUBLIC
17:15IP ADDRESS IS THE SAME AS THE PRIVATE
17:18IP ADDRESS, AND THIS ADDRESS IS
17:20VERY KEY. ENTERPRISE MOVING THROUGH
17:22PUBLIC POINTS AND REALLY LIMITING
17:25A BIG ATTACK FACTOR. NOW GOING BACK
17:30TO THE COMPUTE INSTANCES, I WILL
17:33CREATE THROUGH UI WHERE CHOOSE CPU
17:37OR GPU TYPES AND FOR EACH OF THE
17:39VM TYPES YOU CAN LOOK AT THE SPEC.
17:41THERE YOU CAN SEE THINGS LIKE THE
17:43COST ASSOCIATED WITH IT AND IT ONLY
17:45SHOWS YOU THE VM TYPES WHICH YOU
17:47HAVE QUOTA FOR. IT TAKES ALL OF
17:49THAT INTO CONSIDERATION, AND YOU
17:52CAN ENABLE ACCESS THROUGH PRIVATE-PUBLIC
17:55MECHANISM AND THEN [INDISCERNIBLE]
17:58VNET SETTINGS. SO WHAT I AM GOING
18:01TO DO IS THROUGH AN ARM TEMPLATE
18:04I WILL PROVISION OUR COMPUTE INSTANCE
18:06ON BEHALF OF JUMPEI. SO JUMPEI WILL
18:08BE ABLE TO ACCESS IT, BUT I CANNOT
18:10ACCESS IT, AND THIS IS ON TEMPLATE.
18:12THESE PROPERTIES ARE SIMILAR TO
18:13WHAT WERE IN THE UIN. WHAT IS DIFFERENT
18:18HERE IS THIS IS JUMPEI'S [INDISCERNIBLE]
18:21SO LET ME TAKE IT OFF. THIS IS JUST
18:28A SIMPLE DEPLOYMENT. SO WHILE THAT
18:31IS HAPPENING, I HAD ALREADY BROUGHT
18:33IN A COUPLE OF COMPUTE INSTANCES.
18:35SO THIS IS A COMPUTE INSTANCE WE
18:37HAVE ON BEHALF OF JUMPEI. ONLY JUMPEI
18:39WILL BE ABLE TO ACCESS IT, AND WHEN
18:40I TRY TO ACCESS IT, I'M GOING TO
18:43GET AN ERROR. SEE IT SAYS OKAY,
18:45I CANNOT ACCESS IT BECAUSE THIS
18:46IS CREATE ON BEHALF OF. SO HERE
18:48RBAC IS AT WORK AND BECAUSE JUMPEI
18:51HAS THE RBAC ACTION TO ACCESS THAT
18:53PARTICULAR INSTANCE, BUT I DON'T
18:54HAVE THAT. SO I WILL INSTEAD USE
18:56A COMPUTE INSTANCE WHICH I CREATED
18:58FOR MYSELF, SO I WILL USE THIS GPU
19:01BOX WHICH HAS FOUR GPUS, AND WHAT
19:04I WOULD LIKE TO SHOW HERE IS A GPU
19:11DISTRIBUTED -- YOU WOULD RUN --
19:13AS WELL, BUT I WANTED TO SHOW THE
19:15TRAINING GROUND HERE. SO WHAT IS
19:17INTERESTING HERE IS I'M GOING TO
19:19USE THE COMPUTE, THE INSTANCE THAT
19:22ALREADY EXIST, SO CLUSTER HAVE TO
19:23WAIT FOR IT TO GET PROVISIONED AND
19:24THEN SCALE UP. HERE YOU CAN JUST
19:26USE IT RIGHT AWAY. SO IT WILL FIND
19:31THE EXISTING COMPUTE INSTANCE. AND
19:33IT FOUND IT AND SO IT DOESN'T NEED
19:35TO CREATE ANYTHING, AND JUST YOU
19:39CAN START YOUR -- SO CREATE PROJECT
19:43DIRECTORY, CREATING SCRIPT. SO THIS
19:46TRAINING SCRIPT IS GOING OVER THE
19:48-- DATA SERVICES THERE, AND I WILL
19:55CREATE EXPERIMENT AND THEN AN ESTIMATOR
19:57USING 1. 4, AND THE GPUS , AND THEN
20:03I WILL SUBMIT THE JOB. AND THEN
20:13I CAN -- THE RUN DIRECTLY IN MY
20:16JUPYTER LAB SESSION TO THE WIDGET
20:20OR I CAN GO TO THE ML EXPERIMENTS
20:24TAB AND A RUN WILL START SHOWING
20:26UP HERE PRETTY SOON. SEE IT'S RUNNING,
20:31AND ALSO THIS IS A JOB IT HAS SUBMITTED.
20:33SO THIS WILL TAKE A BIT OF TIME.
20:34SO I ALREADY HAD SUBMITTED ANOTHER
20:37JOB, SO THIS IS ONE OF THE JOBS
20:39I HAVE SUBMITTED EARLIER. SAME RUN,
20:42DISTRIBUTED BY TOUCH, RUN USING
20:45HOROBOT, AND YOU CAN SEE HERE . . .
20:53IT IS MP LAUNCHING WITH FOUR PROCESSES,
20:56SO I'M REALIZING ALL MY FOUR GPUS
20:59IN THE SAME JOB, AND THIS IS FOR
21:02INSTANCE ONE OTHER MPI PROCESS LAUNCHED
21:04ON THE GPU. SO IT SAYS, YEAH, THIS
21:07IS AN MPI JOB, THIS IS FIRST ONE,
21:10AND IT IS USING 1. 4 AND GPU IS
21:13ENABLED FOR THIS JOB. AND YEAH,
21:17SO YEAH, I THINK THAT IS IT. SO
21:21THAT RUN IS STILL GOING ON, AND
21:23YEAH, JUMPEI, PLEASE GO AHEAD. >>
21:26OKAY, SO THANK YOU, VIJETA. YEAH,
21:30BACK TO PRESENTATION, VIJETA BRIEFLY
21:33EXPLAINED THE CLUSTER AND SUBMITTING
21:39THE [INDISCERNIBLE] BY GPU FROM
21:41HIS COMPUTER INSTANCE. HE SAID HE
21:44CREATED COMPUTER INSTANCE ON BEHALF
21:45OF JUMPEI. SO YEAH, LET ME CHECK
21:47THE ACCESS. AND AFTER LET ME QUICKLY
21:51SHOW THE HS CLASS ITSELF. LET'S
22:01CLICK JUPYTER, AS VIJETA SAID. THEN
22:04I CAN ACCESS THE JUPYTER NOTEBOOK.
22:07YES, SO I CAN ACCESS THAT, YEP.
22:10AND IF I GO -- AND AFTER THAT, LET
22:14ME QUICKLY EXPLAIN THE DEPLOYMENT
22:15THROUGH AKS. IN THIS CASE I USED
22:20JUPYTER NOTEBOOK TO DEPLOY, BUT
22:21YOU CAN USE CLI OR UI, WHATEVER.
22:25AND IN THIS CASE, SLOW DOWN AND
22:29IN HERE I CREATE PRIVATE -- SORRY,
22:34AKS, SO I CONFIGURE WIDGET INFORMATION
22:36IN HERE, AND I ALSO LOAD BOTS HERE.
22:39I ALSO ENABLED THE [INDISCERNIBLE]
22:43SO IN ORDER TO DEPLOY YOUR MODEL,
22:48AKS NEEDS [INDISCERNIBLE] THAT YOU
22:49ACCESS THAT DEPLOYED MODEL, RIGHT?
22:51AND BY DEFAULT IT PROVISIONED PUBLIC
22:54LOAD BALANCER WHICH HAS PUBLIC ID,
22:56BUT YOU CAN -- WHICH HAS PRIVATE
22:58ID, AND IT CAN BE ACCESSED ONLY
23:03FROM [INDISCERNIBLE] SO IN THIS
23:05CASE I ENABLED, AND AFTER THAT,
23:10AKS IS PROFICIENT. AFTER THAT, THIS
23:13IS A SQUALLING JOB. IF I ALLOW THIS
23:16SECTION, YEAH, 16. 2 MILLISECOND,
23:21SO SQUALLING END POINT IS WORKING
23:23OVER PRIVATE IP. SO, YEAH, THAT'S
23:26THE DEMO OF THE AKS CLUSTER. IF
23:28I GO BACK TO POWERPOINT SLIDE AND
23:31SUMMARIZE THEM UP, SO BASICALLY
23:33WE PROVISIONED EVERYTHING OVER PRIVATE
23:35END POINT, SO FOR WORKSPACE AND
23:41FOR ASSOCIATED [INDISCERNIBLE] AND
23:42THE VPN ACCESS FROM OUR PC OVER
23:45PRIVATE IP, AND WE ALSO PROVISIONED
23:48PRIVATE IP OF COMPUTE CLUSTER AND
23:51TEST THE COMPUTE INSTANCE SCALEABLE,
23:52AND WE ALSO TRY ON BEHALF OF CREATION.
23:56AND LASTLY WE DEPLOYED MODEL TO
24:01AKS WHICH HAS PRIVATE IP. SO THAT'S
24:03THE DEMO. LET'S MOVE ON TO THE ROAD
24:08MAP. SO YEAH IT IS [INDISCERNIBLE]
24:09ALL THAT ROAD MAP, SO IT HAS NOT
24:12EXPLAINED FEATURES, BUT AS YOU CAN
24:14SEE, WORKSPACE PRIVATE LINK GA AND
24:17YOU CAN ALSO SEE THESE SOURCES IN
24:19THE GREEN TIER, SO IT WILL DO GA
24:22IN NEXT QUARTER. AND PRIVATE IP
24:25IS HERE, SO PROBABLY WILL BE STARTED
24:27BETWEEN OCTOBER TO DECEMBER, AND
24:30GA TARGETED IN 2021. SO YEAH, THAT'S
24:34WHERE WE CAP TODAY. AND WE ALSO
24:37HAVE A RECURRENCES SLIDE, AND NOW
24:40IT'S Q&A. VIJETA, CAN YOU? >> YEP.
24:45THANKS, GREAT PRESENTATION. SO ONE
24:48QUESTION I WAS WONDERING IS [INDISCERNIBLE]
24:49THROUGH REGIONS FOR A LONG TIME.
24:51IS THAT STILL THE CASE? >> YEAH,
24:53THANK YOU VERY MUCH FOR YOUR QUESTION.
24:56SO YEAH, AS OF NOW IT IS AVAILABLE
25:01IN THREE REGIONS, EAST AND WEST
25:04AND SOUTH CENTRAL U. S. AS I SAID
25:09BE AVAILABLE IN ALL PUBLIC AZURE
25:12MACHINE LEARNING REGIONS EXCLUDING
25:18CHINA AND [INDISCERNIBLE] GOVERNMENT.
25:18THEY WILL BE ENABLED BY THE END
25:22OF COUNTRY IN TERM. >> OKAY, THANKS.
25:25AND WHAT DO YOU RECOMMEND AS THE
25:27BEST WAY TO MAKE A PRIVATE WORKSPACE.
25:29>> THANK YOU VERY MUCH FOR A GOOD
25:30QUESTION. SO BASICALLY I RECOMMEND
25:36RESOURCE MANAGER USING RESOURCE
25:40MANAGER TEMPLATE TO PROVISION [INDISCERNIBLE]
25:41WORKSPACE. AND BECAUSE YOU KNOW
25:44IT REQUIRES MULTIPLE CONFIGURATION,
25:46ASSOCIATED WITH THESE RESOURCES
25:46OR ET CETERA. SO YOU CAN AUTOMATE
25:50EVERYTHING USING -- MANAGER TEMPLATE.
25:53SO YEAH, PLEASE REFER TO THE SLIDE.
25:56IT HAS A DIRECT LINK TO THE RESOURCE
25:58MANAGER TEMPLATE. AND WE ALSO SUPPORT
26:01THE DEPLOYMENT FOR AZURE PORTAL.
26:03FOR THE SIMPLER CASE YOU CAN USE
26:06AZURE PORTAL TO DEPLOY AZURE-ENABLED
26:09WORKSPACE. >> OKAY, GREAT. SO THE
26:11NOTEBOOKS, DO THEY WORK WITH THE
26:15WORKSPACE? >> YEAH, THANK YOU VERY
26:15MUCH, YEAH, GOOD QUESTION. SO IN
26:20THE [INDISCERNIBLE] UI SERVICES
26:23DOES NOT WORK WITH [INDISCERNIBLE]
26:24WORKSPACE. NOW ALL SERVICE WE WORK
26:29WITH WORKSPACE EXCEPT INTEGRATED
26:31NOTEBOOK, AND IT LISTEN ENABLED
26:34BY THE END -- WILL BE ENABLED BY
26:36THE END OF JUNE. HEY, I ALSO HAVE
26:39ONE MORE QUESTION, VIJETA. I REMEMBER
26:44MOST OF -- HAS OLD SHUTDOWN FEATURE,
26:47RIGHT? HOW ABOUT COMPUTE INSTANCE?
26:51COMPUTE INSTANCE HAS SHUTDOWN FEATURE?
26:53>> THAT'S A GREAT QUESTION. WE HAVE
26:54HEARD IT FROM OUR CUSTOMERS AND
26:56WE HAVE TAKEN THAT FEEDBACK AND
26:58WE ARE ACTIVELY LOOKING INTO WAYS
27:01TO ADDRESS THIS AND PROVIDE THE
27:05AUTOSHUTDOWN AND AUTOSTART CAPABILITIES.
27:09WE ARE LOOKING INTO THAT RIGHT NOW.
27:12>> AND DO WE HAVE ANY WORKAROUND
27:14FUNCTION? >> RIGHT NOW BECAUSE YOU
27:16CAN DO THE CONTROL OPERATIONS LIKE
27:19STOP AND START, SO YOU CAN SCRIPT
27:21IT, AND USING AS YOUR FUNCTIONS,
27:24YOU CAN TRIGGER IT, BUT WE WERE
27:28GOING TO MAKE IT MUCH EASIER FOR
27:29YOU. >> THANK YOU, VIJETA. IN THIS
27:36PRESENTATION WE EXPLAINED HOW TO
27:40BE SECURE AND COMPLIANT IN MACHINE
27:40LEARNING PLATFORM, AND I HOPE YOU
27:43CAN TRY IT. THANK YOU VERY MUCH
27:44AND PLEASE ENJOY THE REST OF MICROSOFT
27:47IGNITE.