I-Warner Bros. Ukutholwa kufinyelela ukulondolozwa kwezindleko ezingama-60% kanye nokushesha kwe-ML nge-AWS Graviton

Lokhu okuthunyelwe kubhalwe nguNul Sharma, Umphathi Wokufunda Ubunjiniyela, kanye noKarthik Dasani, unjiniyela Wokufunda Umshini Wokufunda, eWarner Bros. Discovery.

I-Warner Bros. Discovery (WBD) iyinkampani ehamba phambili ye-Global Media and Entertainment Company eyakha futhi isabalalisa iphothifoliyo ehlukaniswe kakhulu emhlabeni futhi ephelele yokuqukethwe kanye nemikhiqizo kulo lonke ithelevishini, ifilimu nokusakaza. Ngemikhiqizo yesithonjana kufaka phakathi i-HBO, isiteshi sokutholwa, i-Warner Bros., CNN, DC Ezokuzijabulisa, kanye nabanye abaningi, i-WBD ihambisa izindaba ze-premium ezinhlanganweni emhlabeni jikelele kanye namava. Our streaming services, including HBO Max and discovery+, represent a cornerstone of our direct-to-consumer strategy, offering viewers unprecedented access to our 200,000+ hours of programming.

Kulokhu okuthunyelwe, sichaza isilinganiso seminikelo yethu, izidingo ze-Artificial Intelligence (AI) / Umshini Wokufunda Ingqalasizinda Yezingqalasizinda ze-Amazon Rendestractor ai Izimo zemisebenzi yethu ye-Ml

Emhlabeni ovela ngokushesha wokuzijabulisa wedijithali, okuqukethwe okuhlukile kukodwa akunawona-izibukeli ezidingayo ukuthola izinhlelo ezihambelana nezintshisekelo zazo ezihlukile. Ukuletha okuqukethwe okwenziwe umuntu uqobo sekubalulekile ekubandakanyeni izithameli, izikhathi zokubuka zokushayela, nokwakha ubudlelwano obuhlale nabasebenzisi. Ukuze ukhole ngempumelelo isisekelo sethu esihlukahlukene sabasebenzisi abangaphezu kwe-125m + emazweni angama-100 + (kusukela ngo-2025), sisebenzisa isayensi yedatha, ukuhlaziya kokuziphatha komsebenzisi, kanye nokuhlaziywa komuntu ukubikezela ukuthi yini ababukeli abazoyithanda. Umsebenzi wethu ugxile ekuncomeni ama-algorithms ashukumisayo ama-algorithms ashukumisayo kanye neziphakamiso zokuvumelanisa kokuthandwayo ngakunye, ngenkathi uhlola ngokuqhubekayo futhi ucoca amasu wokuthuthukisa ukunemba kokunye ukunemba okuqukethwe.

Inselelo: Ukulinganisa ukwenziwa komuntu uqobo emhlabeni jikelele ngenkathi kuphathwa izindleko

Ukusesha kwengqalasizinda ye-HBO Max kanye nokwenza izinto ngezinto ezisetshenziselwa izifunda eziyi-9 ze-AWS kwasakazwa ngaphesheya kwe-USA, i-Emea kanye ne-APAC ukuletha izincomo ezenziwe zasendaweni ezenzelwe izintandokazi zesifunda. Le ngqalasizinda ebanzi isenza sikwazi ukugcina izidingo ezingaguquki ezingaguquki ze-100 100ms ngenkathi zisebenzela izincomo zokuqukethwe ezenziwe ngezifiso ezifundeni ezahlukahlukene.

Ngephothifoliyo enkulu yemikhiqizo yethu ethandekayo ehlanganiswe nesisekelo sabasebenzisi esihlukahlukene, sabhekana nenselelo yokwenza izincomo zokuqukethwe ngaphandle kokuyekethisa kwisabelomali. Izinhlelo zokuncoma ziyi-LatCery Clignacy; Badinga ukugijima ngesikhathi sangempela okusho ukuthi izidingo eziqinile zengqalasizinda ye-ML ezidingekayo ekuthumbeni izinsizakalo zethu. Le nselelo yokutholwa kokuqukethwe idinga amasistimu ampchististiteated active angenza ngokuthembekile ngezinga elikhulu, ngisho nangezikhathi ezinkulu zethrafikhi lapho kufika amathrafikhi afinyelela kuma-500% kungakapheli imizuzu engama-500. Besifuna ukusebenza kwesikhathi sangempela kanye nesixazululo sezingqalasizinda esibizayo sezindleko zomsebenzi wethu we-AI / ML.

Isixazululo sethu: Kusetshenziswa ama-AWS Graviton ukuthola ukungasebenzi kahle kwe-ML

Isixazululo sethu sihlanganise ubuchwepheshe obukhulu be-AWS: AWS Graviton processors kanye ne-Amazon Sagemaker AI. Le ndlela ehlanganisiwe yasivumela ukuba sibhekane nekheli ngokusebenza kwethu kanye nezinselelo zezindleko.

Ama-AWS Graviton ngumndeni wabaprosesa owenzelwe ukuletha ukusebenza kwamanani amahle kakhulu wokulayisha amafu asebenza e-Amazon Elastic Accoute Cloud (ama-Amazon EC2) nezinsizakalo eziphethwe ngokuphelele. Baphinde balungiselelwe ama-ML bacloads, kufaka phakathi izinjini zokucubungula ze-Neon Vector, ukusekelwa kweBloaty16, ukunwetshwa kwe-vector (SCOSTIX (SLASTIX (SMLA), kubenza ukuba kube yikhetho oluhle kwizinhlelo zethu zokubonisa ubucayi be-Latency Card.

Sinqume ukuzazama amamodeli wethu we-XGboost kanye namamodeli we-tensorflow asekelwe kulo inqubo yezinyathelo ezimbili. Okokuqala, saqala ngemvelo ye-sandbox, izisebenzi ezihleliwe ezinhle kanye nemicu yokwandisa ukugcwaliswa kwesibonelo esisodwa futhi yabona ukusebenza okungcono kakhulu ngokuqhathaniswa nezimo ezisuselwa ku-x86 emikhunjini yethu. Okwesibili, sathuthela eTraffic Straffic, lapho senza ukuhlolwa kwesithunzi ukuqinisekisa izindleko nezinzuzo zokusebenza esizibonile endaweni yokuma. Siphawule ukuthi izimo ze-graviton zakwazi ukukala cishe ngokuqondile ngisho naku-CPU ephakeme. Siphinde sivule ama-auto-scale silungiselela ukukhulisa isisetshenziswa sokusetshenziswa kwesimo kanye ngoba izimo ze-graviton zakwazi ukubhekana nokuqhuma ithrafikhi ngempumelelo, sanciphisa futhi inani elincane lezimo. Ngokwengeziwe, silinganise ukusebenza kwezindleko vs ukuthi ungamtholeli umuntu ekwenzeni imali ngokweqile komunye.

Umthengiseli we-sagemaker we-wefence wadlala indima ebalulekile ekutholeni ukuhamba kwethu komsebenzi wokuhlola. Ngokushintsha inqubo yokulinganisa uphawu ezinhlotsheni zezezimali ezahlukahlukene kanye nokucushwa, leli thuluzi lincishiswe kakhulu isikhathi esidingekayo ukukhomba ukusetha okuphelele kwamamodeli ethu. Ukuhlaziywa kokusebenza okuzenzakalelayo kusisize ukwenza izinqumo eziqhutshwa idatha mayelana nokukhethwa kwesimo futhi kusheshise ipayipi lethu lokuhambisa imodeli.

Ukuze uqinisekise ukusebenza kanye nokwethenjwa kwengqalasizinda yethu entsha, sisebenzise amakhono okuhlola isithunzi we-Amazon Sagemaker. Lolu hlaka lokuhlola lunikeze amandla iqembu lethu ukuthi lihlaziye ukuthunyelwa okusha eceleni kwezinhlelo ezikhona zomkhiqizo, ukuhlinzeka ngokuqhathanisa nokusebenza kwangempela komhlaba ngaphandle kokufaka umthelela kubasebenzisi bethu. Le ndlela ifakazele ikakhulukazi abasebenzisi beqembu lethu lokufunda lomshini (MLP) njengoba behlola ukuguqulwa kwengqalasizinda ehlukahlukene. Ngokusebenzisa izivivinyo ezihambisanayo zama-setups ahlukene we-Hardware kanye nama-parameter okulandela amahle, singahlola kahle ukusebenza kohlelo ngaphambi kokuzishintsha. Le ndlela yokuhlola amasu isisizile ukuthi silindele izingqinamba ezingaba khona futhi zisebenzise ukucushwa kusenesikhathi kwinqubo yethu yokuthumela.

Umdwebo olandelayo uqokomisa ukuthunyelwa kokuphela kokuphela komsebenzi wethu we-ML woku-afence kuma-AWS. Njengoba kukhonjisiwe lapha, sesivele sisebenzisa izinsizakalo eziningi ze-AWS eziphathwe ngokugcwele njenge-Amazon SageMaker, i-Amazon Isitoreji Service (I-Amazon S3), kanye ne-Amazon DynamoDB ukufezekisa izinhloso zethu ze-Rectunder. Kulokhu, sasithatha isinyathelo esisodwa saya phambili ukufudukela ku-AWS Graviton-based Stances eholele ekugcinweni kwezindleko kanye nokwenza ngcono ukusebenza.

Umphumela

Ukufuduka ku-AWS kususelwa ku-Graviton ezikhungweni ezivela ezimweni ze-X86-Base zilethe imiphumela emangalisayo kulo lonke iphothifoliyo yethu yohlelo lokuncoma.

Kufinyelelwe imali engu-60% yokonga

Ukuhlaziywa kwethu okuphelele kuveze ukuncishiswa kwezindleko ezinkulu kuwo wonke amamodeli ethu okuzibonela, ukufeza imali ephakathi yokonga ngo-60%. Ukuthuthuka bekuphawuleka ikakhulukazi kumamodeli wethu wekhathalogi, lapho sabona khona ukuncishiswa kwezindleko okufika ku-88%.

Izilinganiso ezithuthukile ze-average ne-P99, ezisukela ku-7% ziye ku-60% amamodeli ahlukene

Ngaphandle kokonga izindleko, siphinde sathola izithuthukisi zokusebenza ezibalulekile. Ukuthuthuka kwe-latency ye-P99 bekuhlaba umxhwele kulo lonke imodeli yethu yemodeli ngemodeli yethu ye-XGBOOST ekhombisa ukuncishiswa okumangalisayo kwama-60% ekunciphiseni i-latency. Amanye amamodeli kuphothifoliyo yethu ebonise ukuthuthukiswa okungaguquki okungahambela kuze kube yi-21%. Ideshibhodi elandelayo kusuka ekuhlolweni kwe-A / B Ukuhlola okuvela ku-AWS kususelwa ku-AWS Graviton-based ML Izimo ze-AWS kuthuthukise ama-average kanye nama-P99 latencies bese usika isibonelo ukubala kakhulu. -Yimpatho Imigqa ivela kumaseva we-x86 asekwe emiflombe yethu kanye -Qanda Imigqa isuka kumaseva we-AWS Graviton esekwe.

Isipiliyoni somsebenzisi esithuthukisiwe

Ngokunciphisa i-latency, sathuthukisa kakhulu ukusebenza kwezinsizakalo zethu kanye nesipiliyoni somsebenzisi kumakhasimende ethu; Ababukeli bathola izincomo ezisabelayo ezihambisana kangcono nezintshisekelo zabo.

Uthole ukufuduka komthungo

Sibe nokubambisana okuhle ne-AWS Akhawunti kanye namaqembu wesevisi kulo lonke iphrojekthi. Ukufuduka kwakungenamthungo. Kusukela ekubhekeni kokuqala kokufuduka kokugcina kwathatha inyanga eyodwa; Ubufakazi bomqondo kwimodeli yokulinganisa ikhathalogu enikeze ukulondolozwa kwezindleko okungu-60% kwenziwa ngesikhathi esingangesonto, okwakushesha kakhulu kunesikhathi esasikulinganiselwe ekuqaleni.

Eshukumiseleke ukufeza i-100% yohlelo lwe-repunder ukuze lisebenze ngokusekelwe eGraviton

Uma ubheka imali ekhokhwayo etholakele esikutholile ngohlobo lwe-Graviton Adoptint, njengamanje sisebenzela ukufuduka amamodeli ethu asele kwi-Graviton ngelitshe lokuthola ama-100% wohlelo olususelwa eGraviton.

Ukugcina

Ngokufudukela imithwalo yethu ye-ML yokungena ku-AWS ku-AWS Graviton-based States, siguqule ukuthi siletha kanjani izincomo zokuqukethwe ezenzelwe wena uqobo kubasebenzisi bethu abayi-125m + emazweni angu-100 +. Ukufuduka kwaveza imiphumela ehlaba umxhwele ngokuncishiswa kwezindleko kufinyelela kuma-60% ezinsukwini zethu ze-repunder kanye nentuthuko ye-latency esukela ngo-7% kuya ku-60% amamodeli ahlukene. Lezi zinzuzo zokusebenza zihunyushelwe emiphumeleni yebhizinisi ebonakalayo: Izibukeli zithola izincomo eziphendulayo ezihambelana kangcono nezintshisekelo zazo, okuholele ekuzibayeni okujulile, nokugcinwa okuthuthukile ezinhlelweni zethu, konke ngenkathi kusinika amandla okusebenza kwethu kahle.

Sekukonke, ukwamukelwa kwama-AWS Graviton processors kusibonelo ukuthi izixazululo zefu ezintsha zingenza kanjani ukusebenza kahle kwezezimali nenani lebhizinisi. Isipiliyoni sethu sibonisa ukuthi izinhlangano zingalinganisa ngempumelelo izimfuno zokuncintisana zokusebenza, izindleko, kanye nesilinganiso ekuvezeni ibhizinisi elivela ngokushesha. Njengoba siqhubeka nokwandisa ingqalasizinda yethu ye-ML, lokhu kuthuthuka kuzosisiza ukuthi sihlale sincintisana ngenkathi siletha okuhlangenwe nakho okwenzeka ngokwezifiso ngokwezifiso ezilalelini zethu zomhlaba.

Ukufunda okwengeziwe, bhekisa kokulandelayo:

Iqembu le-WBD lingathanda ukubonga uSunita Nadampalli, u-Utsav Joshi, Karthik Rengasamy, Tito Panicker, Sapna Patel, ne-Gatel Patel, kanye ne-Gatel Panth kusuka kuma-AWS abo kulesixazululo.

Mayelana nababhali

Nulul Sharma ngumshini wokufunda ubunjiniyela bomshini oneminyaka engu-18 + yesipiliyoni ehola amaqembu wobunjiniyela abasebenza kakhulu namaqembu e-MLops e-Warner Bros. Discovery. Amakhono ekwakheni izixazululo ezikali, amaphayiphu we-end-to-end ml, amasistimu wamafu, ne-CD / CD. Irekhodi elifakazelwe ekuletheni okuthobekayo kanye nezisombululo ze-MLOPS eziqhuba ukusebenza kahle nokukhula.

Karthik Dasani Ingabe unjiniyela womshini womshini wokufunda onobuchwepheshe ezinhlelweni zokuncoma ezikhulu kanye nama-ML ops e-Warner Bros. Discovery. Unolwazi olunzulu ekukhiqizeni izixazululo ze-AI ngokugxila okuqinile ekusebenzeni nasekulahlekeni kahle. Umsebenzi wakhe wamabhuloho asetshenzisiwe ucwaningo nezinhlelo zokufunda zomhlaba jikelele.

Mayelana neqembu le-AWS

USunita Nadampalli Ungunjiniyela ophambili kanye nochwepheshe we-AI / ML kuma-AWS. Uhola ukusebenza kokusebenza kwesoftware ye-AWS Graviton Software ye-AI / ML kanye nemithwalo ye-HPC. Unothando ngokuthuthukiswa kwesoftware evulekile kanye nokuletha izisombululo zesoftware eziphakeme nezinzile ze-SOCS ezisuselwa ku-ARM ISA.

Utsav Joshi ngumphathi we-akhawunti oyinhloko yezobuchwepheshe ngama-AWS. Uhlala eNew Jersey futhi ujabulele ukusebenza ngamakhasimende ama-AWS ekuxazululeni izinselelo zokwakha, ukusebenza, kanye nezinselelo zokusebenzisa imali. Esikhathini sakhe sokuphumula, uyakujabulela ukuhamba, uhambo lomgwaqo, futhi edlala nezingane zakhe.

UKarthik Rengasamy Ingabe ukwakhiwa kwezixazululo eziphezulu kuma-AWS, abhekelele ekusizeni amakhasimende e-Media and Entertainment assign Ugxile ku-Media Supply Chain, okulobongo, kanye nezixazululo zokusakazwa kwevidiyo, ukusebenza ngokusondelene namakhasimende ukushayela izinto ezintsha futhi wandisa imithwalo yomsebenzi wemidiya kuma-AWS. Uthando lwakhe lulele ekwakheni izixazululo eziphephile, ezinambithayo, nezingabizi kakhulu eziguqula indlela imithombo yezindaba ilawulwa ngayo futhi ilethwe kubuthameli bomhlaba.

I-Tito Panicker I-SR. I-Global Solutions Solutions Solutions esiza amakhasimende amakhulu ebhizinisi aphephile, acwebezelayo, kanye nezisombululo ezikhuthazayo efwini. Indawo yakhe eyinhloko yokugxila yimithombo yezindaba nokuzijabulisa mpo, lapho ihamba phambili ekusakazweni okuqondile (D2C) kwabathengi (idatha / ama-analytics, i-AI / ML, kanye ne-ai / ml ai.

I-Sapna Patel Ingabe Umphathi Wezixazululo Eziyinhloko Zezixazululo ezi-AWS ezisiza amakhasimende e-Media and Entertainment enze ngcono uhambo lwazo lwefu ngokuholwa ngamasu kanye nokuphathwa kobudlelwano. Ugxile ekuqhubeni impumelelo yamakhasimende ngokuvumelanisa izixazululo zezempi nezinhloso zebhizinisi, okwenza ukuthi amakhasimende andisa inani kusuka ekutshalweni kwemali kwamafu ngenkathi efinyelela izinhloso zawo zezobuchwepheshe nezisebenzayo.

I-Gautham Panth Ungumphathi womkhiqizo oyinhloko kuma-AST wagxila ekwakheni izixazululo zengqalasizinda yamafu. Eneminyaka engaphezu kwengu-20 yobuchwepheshe bokuqondiswa kwezigwegwe okuthatha i-computing yamafu, ingqalasizinda yebhizinisi, nesoftware, i-Gautham ithola ukuqonda kwakhe okuphelele kwezinselelo zamakhasimende, ukushayela ukuqondiswa okuzayo kanye namakhono weminikelo ye-AWS.