Chii chaitika
Vatsvagiri vakaunza DataKernelBench, bhenji rekuyedza kana mamodheru emitauro mikuru anogona kukwirisa zvisina kujairika, data-kufamba-inorema dhatabhesi mashandiro paGPUs. Iyo sisitimu inoshandura mivhunzo yeSQL kuita yakasimbiswa PyTorch TorchPlan zvirongwa, yobva yaongorora mamodheru sezvavanokwirisa ingave yepakati tensor-yakasungwa kodhi chikamu kana kuzere kubvunza muCUDA kana Triton.
Iro bepa, rakatumirwa kuarXiv muna Aug. 25, 2026 uye rakaonekwa munzvimbo yakagamuchirwa paEMNLP 2026, rinopa DataKernelBench seongororo yakanangana neAI-yakagadzirwa optimization yedatabase mibvunzo paGPUs. Vanyori vanopokana kuti iripo LLM kernel mabhenji haaedze zvakakwana dhatabhesi-maitiro evashandisi, ayo anogona kuve asina kujairika, akasiyana uye anotongwa nekufamba kwedata. Izvo zvinoita kuti chinangwa chebhenji chisiyane kubva kune vanogaro shanda vanowanzo shandiswa kuongorora yakagadzirwa GPU kodhi.
DataKernelBench inoshandura SQL kuita yakasimbiswa PyTorch TorchPlan zvirongwa. Iwo mamodheru anozokumbirwa kukwidziridza ingave yepakati tensor-yakasungwa snippet kana iyo yese mubvunzo, uchishandisa CUDA kana Triton. Ongororo iyi inosanganisira kuuraya-inotungamirwa kugadzirisa, zvichireva kuti zvirongwa zvinogadzirwa zvinoedzwa uye zvinodzokororwa kuburikidza nemhinduro kubva mukuitwa kwavo. Chidimbu chinoti chidzidzo ichi chinobata gumi evaridzi uye akavhurika huremu modhi paTPC-H SF10 basa rekushandisa uchishandisa H100 GPU.
Zvinoenderana nemhedzisiro yebepa yakashumwa, iyo yakasimba-yakazara-mubvunzo CUDA kumisikidzwa yakawana 2.11 × yekumhanyisa pamusoro peiyo yekutanga yekuenzanisa pachiyero chekupasa. Kwakabva hakuburitse zita rekutanga nekuti chidimbu chine chinongedzo chisina kurongeka munzvimbo iyoyo, saka iyo chaiyo yereferensi poindi haigone kuzivikanwa kubva pane yakapihwa. Vanyori vanoshumawo kuwedzera-chikuru chekuwedzera: TorchPlan yakasanganiswa neDask-cuDF yedata yakakura kupfuura GPU ndangariro, uye paTPC-H SF100 vachishandisa mana H100 GPUs, sisitimu yakawana yakashumwa 2.54 × kumhanya.
Kutorwa pamwe chete, iyo setup inotsanangura nhevedzano kubva kumubvunzo unomiririra kune inogadzirwa chirongwa chekuita uye kuita kwakashumwa. Iyo SQL yekupinda inomiririrwa seyakasimbiswa TorchPlan chirongwa, nepo optimization tarisiro inogona kunge iri yepakati tensor-yakasungwa chikamu kana yakazara mubvunzo. Mhedzisiro inogadzirwa nemuenzaniso haina kubatwa seyakakwana chete nekuti yakanyorwa; kuongororwa kunoshandisa kuuraya-inotungamirwa kugadzirisa kuyedza uye kugadzirisa zvirongwa kuburikidza nemhinduro kubva mukuurayiwa. Iyo hardware uye magadzirirwo ebasa zvakare chikamu cheyedzo yakashumwa: iyo abstract inotsanangura gumi evaridzi uye akavhurika huremu modhi, iyo TPC-H SF10 basa rekuita, uye H100 GPU. Iyo yakasiyana inotsanangura yakakura-mwero TorchPlan uye Dask-cuDF kumisikidzwa paTPC-H SF100 ine mana H100 GPUs. Mukati meiyo chimiro, iyo yakashumwa yekumhanyisa mhedzisiro yezvakatarwa zvigadziriso uye mamiriro ekupfuura, nepo iyo isina kurongeka yekutanga link inosiya yekufananidza referensi isina kutaurwa mune yakapihwa sosi. Iyi ndiyo chiyero chezvinopihwa neabstract nezve bhenji uye ongororo yayo yakashumwa. Tsananguro yacho inoburitsa zvinhu zviri kuvandudzwa, sarudzo dzehurongwa, magadzirirwo ekugadzirisa, basa rakaongororwa, zvigadziriso zvehardware, uye zviviri zvakashumwa maitiro ekuita, asi hazviwedzeri zvakawanda kupfuura izvo zvinopihwa muchidimbu.
Nei zvichikosha
Iro basa rinogadzirisa gaka mune iripo LLM coding mabhenji, ayo vanyori vanoti zvakanyanya kutarisisa kune-mashini ekudzidza-vashandisi kwete mabasa edatabase. Kana mibairo yakashumwa ikabata mabasa akawandisa, mhando dzemitauro dzinogona kuve vabatsiri vanobatsira kune nyanzvi yekernel engineering inodiwa kuita GPU-inomhanyisa dhatabhesi nekukurumidza.
Izvo zvakakosha ndezvekuti kuita kwedhatabhesi kunowanzoenderana nekuremerwa kwebasa-chaiwo sarudzo dzekuita, kwete chete pakusarudza inokurumidza general-chinangwa system. Chirevo chepakati chebepa ndechekuti maLLM anogona kutora chikamu mune iyi yakasarudzika optimization maitiro nekugadzira uye kugadzirisa zvirongwa zveGPU zvemibvunzo yakazara. Izvo zvinoisa modhi padhuze nechimiro cheiyo chaiyo dhatabhesi basa rekuita pane bhenji inoongorora ega ega emuchina-yekudzidza kernels.
Izvo zvakashumwa zvakawanikwa zvinoratidzawo kupatsanurwa kwevashandi pakati pekugona kwemuenzaniso uye ruzivo rwebasa. Vanyori vanoti kuita kwepamusoro-kuita kunowanzo shandisa kernel fusion uye shanduko pakuita zano. Vanoshuma zvakare kuti mamiriro ebasa akakosha kupfuura mamiriro ehardware, uye kuti mamodheru akasimba anobatsira zvakanyanya kubva kuzere-mubvunzo nyanzvi. Mukutaura, mhedzisiro inoratidza kuti kupa iyo modhi ine tsananguro yakadzama yemubvunzo uye data rayo zvingave nebasa kupfuura kungotsanangura iyo GPU iyo kodhi ichamhanya.
Mhedzisiro yacho yakakosha segwara rekutsvagisa, asi hausi humbowo hwekuti injiniya yedatabase yave otomatiki mukugadzira. Sosi yacho inotsanangura bhenji uye inodzorwa zviedzo, kwete kutumirwa mune mhenyu dhatabhesi sevhisi. Izvo zvakare hazviratidze kuti iyo kodhi inogadzirwa ndeyechokwadi kunze kweyakaedzwa mibvunzo, kuti zvakachipa kugadzira kupfuura zvakanyorwa nevanhu, kana kuti kumhanya kwaizopona kuchinja kugoverwa kwedata uye mashandiro ezvinodiwa. Iyo miganho yakakosha nekuti inokurumidza kubvunza iyo inotadza pamupendero kesi haisi inoshandisika dhatabhesi optimization.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mibvunzo mikuru ndeyekuti mhedzisiro yacho inowanda kupfuura iyo TPC-H yebasa yakayedzwa, kuti bhenji rinotsanangura sei kupasa kwakazara, uye kana iyo yakashumwa yekumhanyisa inoramba iripo mushure me accounting yekusimudzira, kusimbiswa uye mutengo wehardware. Iyo yakapihwa sosi ndeye abstract, saka izvo ruzivo uye yakazvimirira kudzokorora zvinoramba zvisina kugadziriswa.
Nyaya yekutanga kutarisa ndeye reproducibility. Iyo abstract inoratidza nhamba uye yakakura mhando dzemodheru, iyo Hardware, uye TPC-H chiyero zvinhu, asi haitauri mamodheru, inotsanangura kukurumidza kwavo, tsanangura iyo pass-rate kuverenga, kana kupa iyo yakazara yekufananidza yekutanga. Iwo mameseji ndiwo achaona kuti masisitimu akaongororwa zvakadii uye kuti vamwe vanogona sei kudzokorora zviedzo.
mumwe mubvunzo wakavhurika. Miedzo yakashumwa inoshandisa TPC-H SF10 pane imwe H100 GPU uye TPC-H SF100 pane ina H100 GPUs. Iyo sosi haitaure kana DataKernelBench inovhara dzimwe mhuri dzemibvunzo, injini dze database, kugovera data, zvizvarwa zveGPU kana kusanganiswa kweCPU-GPU deployments. Izvo zvakare hazvitaridzi kuti maitiro anoita sei kana iyo data ichipfuura ndangariro nenzira dzisina kurongeka, kana kana iko kurongeka uye latency inofanirwa kuchengetedzwa pasi pekuchinja mitoro yekugadzira.
Chekupedzisira, ongororo dzenguva yemberi dzinofanira kuparadzanisa kumhanya kwekuuraya kubva pamutengo wakazara wekushandisa LLM-based optimizer. Matanho akakodzera anosanganisira nguva yechizvarwa, nhamba yekuedza kugadzirisa, kusimbiswa kwepamusoro, kushandiswa kweGPU, kushandisa ndangariro uye mutengo wekukundikana kana kurambwa zvirongwa. Sosi inoti kuuraya-inotungamirwa kugadzirisa chikamu cheiyo nzira, asi iyo abstract haipe manhamba eiyo maitiro. Kusvikira izvo zvisingazivikanwi zvaziviswa uye zvakaedzwa zvakazvimirira, iyo 2.11 × uye 2.54 × nhamba dzinofanirwa kubatwa semibairo inotaurwa nebepa rino pasi pemamiriro ayo ekuedza akataurwa, kwete sechivimbiso chekuita kwese.