Tsarin Sadarwar Ƙarfin Lissafi Mai Ƙayyadaddun Aiki: Tsari, Fasahohi, da Hasashe

1. Gabatarwa

Saurin ci gaban aikace-aikace kamar Hankalin Wucin Gadi (AI), tuƙi mai cin gashin kansa, Gaskiyar Gaskiya ta Girgije (VR), da masana'antu masu hankali sun haifar da buƙata da ba a taɓa ganin irinta ba ga cibiyoyin sadarwa waɗanda ba kawai ke ba da garantin babban bandeji ba, amma aikin da ya ƙayyade a cikin jinkirin watsawa da aiwatar da lissafi. Cibiyoyin sadarwa na "Mafi Kyawun Ƙoƙari" na gargajiya da sarrafa albarkatun lissafi keɓaɓɓu ba su isa ba. Wannan takarda ta gabatar da Tsarin Sadarwar Ƙarfin Lissafi Mai Ƙayyadaddun Aiki (Det-CPN), sabon tsari wanda ya haɗa ƙa'idodin sadarwa masu ƙayyadaddun aiki da tsara ƙarfin lissafi don samar da ayyuka masu garantin ƙarshe-zuwa-ƙarshe ga ayyuka masu saurin lokaci da na lissafi mai tsanani.

2. Bayanan Bincike da Dalili

2.1 Tashin Aikace-aikacen Lissafi Mai Tsanani

Aikace-aikacen zamani suna da fuska biyu: duka biyun suna da saurin gaggawa kuma lissafi mai tsanani. Misali, ƙididdiga na ainihi don tuƙi mai cin gashin kansa dole ne ya sarrafa bayanan firikwensin cikin ƙayyadaddun ƙayyadaddun lokaci, yayin da VR na girgije ke buƙatar yin zane-zane masu rikitarwa tare da ƙaramin jinkiri daga motsi zuwa photon. Wannan yana haifar da "ratar ƙayyadaddun aiki" inda ba sadarwar ƙarfin lissafi (CPN) ko sadarwa mai ƙayyadaddun aiki (DetNet) kaɗai ba za su iya samar da cikakkiyar mafita ba.

2.2 Iyakokin Tsarin Na Yanzu

Binciken CPN na yanzu yana mai da hankali kan ingantaccen tsara aikin lissafi amma sau da yawa yana ɗaukar cibiyar sadarwa a matsayin akwatin baƙi tare da jinkiri mai canzawa. Akasin haka, DetNet yana tabbatar da iyakancewa, ƙananan jinkirin isar da fakiti amma baya la'akari da ƙayyadaddun lokacin aiwatar da ayyukan lissafi da kansu a wurin ƙarshe. Wannan hanyar da aka raba ta kasa aikace-aikacen da ke buƙatar garantin cikakken lokacin kammalawa daga ƙaddamar da aiki zuwa isar da sakamako.

3. Tsarin Sadarwar Ƙarfin Lissafi Mai Ƙayyadaddun Aiki (Det-CPN)

3.1 Bayyani Gabaɗaya na Tsarin Tsarin

Tsarin Det-CPN da aka tsara tsarin tsari ne mai yawan matakai wanda aka tsara don sarrafawa ɗaya. Ya haɗa:

• Matakin Aikace-aikace: Yana ɗaukar nauyin ayyuka masu saurin gaggawa da na lissafi mai tsanani.
• Matakin Sarrafawa Gudaɗaya: Kwakwalwar Det-CPN, mai alhakin haɗin tsara albarkatu, sarrafa yanayin duniya, da tsara sabis na ƙayyadaddun aiki.
• Matakin Albarkatu: Ya ƙunshi tushen tsarin sadarwa mai ƙayyadaddun aiki (masu sauyawa, masu karkatar da hanyoyin sadarwa tare da siffa mai sanin lokaci) da nodes na lissafi daban-daban (sabobin gefe, cibiyoyin bayanai na girgije, na'urori na musamman na AI).

Lura: Zane na ra'ayi zai nuna waɗannan matakan tare da kibiyoyi biyu tsakanin Matakin Sarrafawa Gudaɗaya da Matakin Albarkatu, yana mai da hankali kan tsarin tsakiya.

3.2 Babban Ƙarfin Fasaha

Det-CPN yana nufin samar da ginshiƙai huɗu na ƙayyadaddun aiki:

1. Ƙayyadaddun Jinkiri: Garantin iyaka mafi girma akan jinkirin fakiti daga ƙarshe zuwa ƙarshe.
2. Ƙayyadaddun Jinkiri: Garantin iyaka akan bambancin jinkiri (mafi kyau kusa da sifili).
3. Ƙayyadaddun Hanya: Hanyoyin turawa bayanai masu iya faɗi da kwanciyar hankali.
4. Ƙayyadaddun Lissafi: Garantin lokacin aiwatarwa don aikin lissafi akan takamaiman albarkatu.

3.3 Tsarin Aiki na Det-CPN

Tsarin aiki na yau da kullun ya ƙunshi: 1) Mai amfani ya ƙaddamar da aiki tare da buƙatu (misali, "kammala wannan ƙididdiga cikin 20ms"). 2) Mai Sarrafawa Gudaɗaya yana fahimtar hanyoyin sadarwa da albarkatun lissafi da ake da su. 3) Ya haɗa lissafin mafi kyawun hanya da sanya node na lissafi wanda ya dace da ƙayyadaddun ƙayyadaddun. 4) Yana adana albarkatun kuma yana tsara ƙayyadaddun watsawa da aiwatar da lissafi.

4. Muhimman Fasahohin Da Suka Ba Da Damar

4.1 Tsara Sadarwa Mai Ƙayyadaddun Aiki

Yana amfani da fasahohi daga IETF DetNet da IEEE TSN, kamar Siffa Mai Sanin Lokaci (TAS) da Ci gaba da Jere da Turawa (CQF), don ƙirƙirar hanyoyin da aka tsara, marasa tsangwama don gudun muhimman zirga-zirga.

4.2 Fahimtar Ƙarfin Lissafi da Ƙirƙira Samfuri

Yana buƙatar lissafin albarkatun lissafi na ainihi (nau'in CPU/GPU, ƙwaƙwalwar ajiya da ake da ita, kaya na yanzu) kuma, mahimmanci, samfuri don hasashen lokacin aiwatar da aiki. Wannan ya fi rikitarwa fiye da ƙirar jinkirin cibiyar sadarwa saboda bambancin aiki.

4.3 Haɗin Tsara Albarkatun Lissafi-Sadarwa

Kalubalen algorithm na asali. Mai sarrafawa dole ne ya warware matsalar ingantawa mai ƙuntatawa: Rage farashin albarkatu gabaɗaya (ko ƙara amfani) bisa ga: Jinkirin Cibiyar Sadarwa + Lokacin Aiwatar Aiki + Jinkirin Komawar Sakamako ≤ Ƙayyadaddun Lokacin Aikace-aikace.

5. Kalubale da Hanyoyin Gaba

Takardar ta gano kalubale da yawa: rikitarwar ƙirar albarkatu ta fadin yanki, haɓaka sarrafawa ta tsakiya, daidaitawa tsakanin masu siyarwa, da tsaron matakin sarrafawa. Hanyoyin gaba suna nuni zuwa ga amfani da AI/ML don tsara hasashe, haɗawa da cibiyoyin sadarwa na 6G, da faɗaɗawa zuwa ci gaba da lissafi daga na'urorin IoT zuwa girgije.

6. Babban Fahimta & Ra'ayi na Mai Bincike

Babban Fahimta: Det-CPN shine amsar gine-ginen da ba makawa ga lantarkar matakan masana'antu na zahiri. Yana daidai da cibiyar sadarwa ta motsawa daga sarrafa tsarin ƙididdiga zuwa Six Sigma—yana buƙatar ba kawai matsakaicin aiki ba, amma garantin sakamako, ma'auni, da hasashe. Marubutan sun gano daidai cewa ƙimar tana cikin haɗuwa, ba abubuwan da aka haɗa ba. Cibiyar sadarwa mai ƙayyadaddun aiki ba tare da lissafin da za a iya faɗi ba ba shi da amfani ga bututun ƙididdiga na AI, kuma akasin haka.

Tsarin Hankali: Hujjar tana da inganci: buƙatun lissafi masu fashewa (sun ambaci horar da GPT-3 na shekaru 365 na GPU) sun haɗu da ƙayyadaddun iyakoki na jinkiri (daga sarrafa kai da kai na masana'antu) don ƙirƙirar matsalar da ba za a iya warwarewa ba ga gine-ginen da aka ware. Mafita da aka gabatar ta bi hankali—matakin sarrafawa guda ɗaya wanda ke sarrafa duka yankuna guda ɗaya. Wannan yana kwatanta juyin halitta a cikin lissafin girgije daga sarrafa sabobin da cibiyoyin sadarwa daban-daban zuwa software da aka ƙaddara komai.

Ƙarfi & Kurakurai: Ƙarfin takardar shine ƙayyadaddun matsalar sa da hangen nesa na gabaɗaya. Duk da haka, a fili yana da sauƙi akan "yadda". Tsarin gine-ginen da aka gabatar yana da matsayi mai girma, kuma sashin "mahimman fasahohi" yana karantawa kamar jerin abubuwan da ake so fiye da zane. Akwai ƙarancin tattaunawa a kan yarjejeniyar sarrafawa, tsarin rarraba yanayi, ko yadda ake sarrafa yanayin gazawa cikin ƙayyadaddun aiki. Idan aka kwatanta da ƙaƙƙarfan hanyar, tushen lissafi na ayyukan farko kamar takardar CycleGAN (wanda ya gabatar da cikakken tsari, sabon tsari tare da cikakkun ayyukan asara), wannan shawarar Det-CPN tana jin kamar takarda ta matsayi ko ajandar bincike.

Fahimta Mai Aiki: Ga ƴan wasan masana'antu, abin da za a ɗauka shine fara saka hannun jari a cikin kayan aiki da na'urar lantarki. Ba za ku iya tsara abin da ba za ku iya aunawa ba. Gina cikakkun samfuran ainihin lokacin aiwatar da aikin lissafi aiki ne na R&D mara mahimmanci kamar yadda kamfanoni kamar NVIDIA ke yi don GPUs ɗin su. Ga ƙungiyoyin daidaitawa, fifikon ya kamata ya zama ƙayyadaddun buɗaɗɗen APIs don taƙaita albarkatun lissafi da niyya na sabis na ƙayyadaddun aiki, kama da aikin IETF akan samfuran YANG. Gasar mallakar "Matakin Sarrafawa Gudaɗaya" shine inda za a yi yaƙin dandamali na gaba, tsakanin manyan masu girgije, masu siyar da kayan aikin sadarwa, da ƙungiyoyin buɗaɗɗen tushe.

7. Zurfin Fasaha & Ƙirƙirar Lissafi

Ana iya ƙirƙira babbar matsalar tsara a cikin Det-CPN a matsayin ingantawa mai ƙuntatawa. Bari mu ayyana aiki $T_i$ tare da ƙayyadaddun lokaci $D_i$, girman bayanan shigarwa $S_i$, da ayyukan lissafi da ake buƙata $C_i$. Cibiyar sadarwa zane ne $G=(V,E)$ tare da maki $V$ (nodes na lissafi da masu sauyawa) da gefuna $E$ (hanyoyin haɗi). Kowane node na lissafi $v \in V_c \subset V$ yana da ƙarfin lissafi da ake da shi $P_v(t)$ (a cikin FLOPS) da jerin gwano. Kowane hanyar haɗi $e$ yana da bandeji $B_e$ da jinkirin yaduwa $d_e$.

Mai sarrafawa dole ne ya sami node na lissafi $v$ da hanyar cibiyar sadarwa $p$ daga tushe zuwa $v$ da komawa kamar haka:

$$ \underbrace{\sum_{e \in p_{to}} \left( \frac{S_i}{B_e} + d_e \right)}_{\text{Watsawa zuwa Lissafi}} + \underbrace{\frac{C_i}{P_v}}_{\text{Lokacin Aiwatarwa}} + \underbrace{\sum_{e \in p_{back}} \left( \frac{S_{out}}{B_e} + d_e \right)}_{\text{Komawar Sakamako}} \leq D_i $$

Wannan samfuri ne mai sauƙi. Ƙirar gaske dole ne ta yi la'akari da tsara hanyar haɗi ta hanyar TAS (ƙara ƙayyadaddun taga-lokaci), jinkirin jeri a node na lissafi, da bambancin $P_v(t)$ saboda yawan haya. Warware wannan a cikin ainihin lokaci don zuwan ayyuka masu ƙarfi matsala ce ta haɗaɗɗun ingantawa, mai yuwuwa yana buƙatar dabara ko hanyoyin tushen ML, kamar yadda aka nuna a cikin takardar nassoshi zuwa zurfin ƙarfafa koyo [7].

8. Tsarin Bincike & Nazarin Lamari na Ra'ayi

Yanayi: Masana'anta tana amfani da hangen nesa na inji na ainihi don gano lahani akan layin haɗawa mai sauri. Kyamara tana ɗaukar hoto wanda dole ne a sarrafa shi ta hanyar samfurin AI, kuma dole ne a aika da yanke shawara na wucewa/rasa zuwa hannun mutum-mutumi cikin 50ms don ƙi kayan da ba su da kyau.

Tsarin Det-CPN:

1. Ƙaddamar da Aiki: Tsarin kyamara ya ƙaddamar da aiki: "Nazari hoto [bayani], ƙayyadaddun lokaci=50ms."
2. Gano Albarkatu: Mai Sarrafawa Gudaɗaya ya duba:
   • Cibiyar Sadarwa: Raka'o'in jadawalin TSN da ake da su akan cibiyar sadarwar bene na masana'anta.
   • Lissafi: Sabon gefe A (GPU) yana da nisa 10ms, ƙiyasin lokacin ƙididdiga=15ms. Sabon gefe B (CPU) yana da nisa 5ms, ƙiyasin lokacin ƙididdiga=35ms.
3. Yanke Shawara na Haɗin Tsara: Mai sarrafawa ya lissafta jimillar lokutan:
   • Hanya zuwa A (10ms) + Lissafi (15ms) + Komawa (10ms) = 35ms.
   • Hanya zuwa B (5ms) + Lissafi (35ms) + Komawa (5ms) = 45ms.
   Duka biyun sun cika ƙayyadaddun lokaci. Mai sarrafawa na iya zaɓar Sabon A don ƙarancin jinkiri ko Sabon B don adana albarkatun GPU don wasu ayyuka, bisa ga manufa.
4. Tsari & Aiwatarwa: Mai sarrafawa yana adana tazarar lokacin TSN don gudun kyamara-zuwa-sabon A, yana umurci sabon A don ware zaren GPU, kuma yana tsara ƙayyadaddun watsawa da aiwatarwa.

Wannan lamarin ya nuna yadda Det-CPN ke yin ciniki na sananne a fadin yankuna, wanda ba zai yiwu ba tare da masu tsara cibiyar sadarwa da lissafi daban-daban.

9. Hasashen Aikace-aikace & Hanyoyin Gaba

Aikace-aikace Nan da Nan (shekaru 3-5): 'Ya'yan itace masu ƙanƙanta suna cikin yanayi masu sarrafawa, masu ƙima mai girma:

• Masana'antu Masu Hankali & IoT na Masana'antu: Don sarrafa tsari mai rufaffiyar madauki da haɗin gwiwar mutum-mutumi.
• Professional Cloud XR: Don horarwa, kwaikwayo, da haɗin gwiwar nesa inda jinkiri ke haifar da rashin lafiyar na'urar kwaikwayo.
• Tuƙi da Jirage marasa matuka: Inda dole ne a iyakance jinkirin madauki don aminci.

Hanyoyin Gaba & Iyakokin Bincike:

• Matakin Sarrafawa na Asali na AI: Yin amfani da AI mai haifarwa ko samfuran tushe don hasashen yanayin zirga-zirga da buƙatun lissafi, tsara albarkatu a hankali. Bincike daga cibiyoyi kamar MIT's CSAIL akan algorithms masu haɓaka koyo yana da mahimmanci a nan.
• Haɗin Lissafin Ƙididdiga: Yayin da lissafin ƙididdiga ya girma, tsara damar samun na'urori na sarrafa ƙididdiga (QPUs) akan cibiyar sadarwa tare da ƙayyadaddun jinkiri zai zama mahimmanci ga algorithms na ƙididdiga na gargajiya.
• Metaverse Mai Ƙayyadaddun Aiki: Gina duniyoyin kama-da-wane, rabaɗɗun duniyoyin kama-da-wane na buƙatar sabunta yanayi tare da jituwa a fadin miliyoyin ƙungiyoyi—kalubalen Det-CPN mai girma.
• Daidaitawa & Haɗin kai: Nasarar ƙarshe ta dogara ne akan ma'auni waɗanda ke ba da damar kayan aiki daga Cisco, Huawei, NVIDIA, da Intel suyi aiki tare cikin sauƙi a cikin Det-CPN, mai yiwuwa ƙungiyoyi kamar IETF, ETSI, da Gidauniyar Linux suka motsa.

10. Nassoshi

1. Brown, T. B., et al. (2020). Harsunan Samfuran Ƴan Ƙaramin Malami. Ci gaba a cikin Tsarin Bayanai na Neural, 33.
2. IDC. (2022). Jagorar Kashe Kuɗin Hankalin Wucin Gadi na Duniya.
3. IEC/IEEE 60802. Bayanin TSN don Sarrafa Masana'antu.
4. Liu, Y., et al. (2021). Cibiyar Sadarwar Ƙarfin Lissafi: Bincike. IEEE Jaridar Abubuwan Duniya.
5. Finn, N., & Thubert, P. (2016). Tsarin Gine-ginen Sadarwa Mai Ƙayyadaddun Aiki. IETF RFC 8557.
6. Li, H., et al. (2021). Sadarwa Mai Ƙayyadaddun Aiki don Lissafin Gefe. IEEE INFOCOM Workshops.
7. Zhang, H., et al. (2022). Tsara Mai Ƙayyadaddun Aiki na Tushen DRL don Haɗuwar Lissafi da Sadarwa. IEEE Transactions akan Gudanar da Sabis na Cibiyar Sadarwa.
8. Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hotuna-zuwa-Hotuna mara Haɗin gwiwa ta amfani da Cibiyoyin Adawa na Ci gaba da Zagaye. IEEE Taron Duniya akan Kwamfuta ta Kwamfuta (ICCV). [Nassoshi na waje don ƙaƙƙarfan hanyar]
9. MIT Laboratory na Kimiyyar Kwamfuta & Hankalin Wucin Gadi (CSAIL). Bincike akan Algorithms Masu Haɓaka Koyo. https://www.csail.mit.edu [Nassoshi na waje don hanyar gaba]