COLA: Haɗin Kai na Sarrafa Girma ta Atomatik don Ƙananan Ayyuka na Girgije

1. Gabatarwa

Canjin daga tsarin gine-ginen guda ɗaya zuwa ƙananan ayyuka masu sauƙi a cikin aikace-aikacen girgije yana haifar da rikitarwa mai mahimmanci a cikin sarrafa albarkatu. Dole ne masu haɓakawa su yanke shawarar yawan albarkatun lissafi (misali, kwafin akwati, VMs) da za a raba wa kowane ƙaramin sabis. Wannan shawarar tana tasiri sosai ga duka farashin aiki na mai haɓakawa da kuma jinkirin ƙarshe-zuwa-ƙarshe da mai amfani da aikace-aikacen ke fuskanta. Hanyoyin sarrafa girma ta atomatik na gargajiya, kamar Horizontal Pod Autoscaling (HPA), suna daidaita kowane ƙaramin sabis da kansa bisa ma'auni na gida kamar amfani da CPU. Duk da haka, wannan hanyar ba ta da kyau saboda tana yin watsi da yanayin dogaro da juna na ƙananan ayyuka a cikin aikin aikace-aikace. An gabatar da COLA (Collective Autoscaler) a matsayin mafita wanda ke tara rarraba albarkatu a duk ƙananan ayyuka tare da manufa ta duniya: rage farashin kuɗi yayin tabbatar da cewa jinkirin ƙarshe-zuwa-ƙarshe na aikace-aikacen ya kasance ƙasa da wani manufa da aka ƙayyade.

2. Matsalar Sarrafa Girma Mai Zaman Kansa

Madaidaicin ma'auni na masana'antu na sarrafa girma ta atomatik yana aiki ta hanyar rarraba, kowane ƙaramin sabis. Kowane sabis yana haifar da ayyukan daidaitawa (ƙara/cire VMs ko gwangwani) lokacin da amfanin albarkatunsa (CPU, ƙwaƙwalwar ajiya) ya wuce kofa. Babban aibi shi ne cewa wannan ra'ayi na gida ya kasa yin la'akari da aikin duniya na aikace-aikacen. Inganta jinkiri na ƙaramin sabis ɗaya na iya yin tasiri kaɗan akan jimillar jinkirin da mai amfani ya gane idan wani sabis a cikin sarkar ya kasance cikas. Wannan yana haifar da rarraba albarkatu mara inganci—yawan tanadi wasu ayyuka yayin da ake ƙarancin tanadi ga mahimman cikas—wanda ke haifar da farashi mafi girma ba tare da cimma manufar ƙayyadaddun matakin sabis (SLO) na jinkiri da ake so ba.

3. COLA: Hanyar Haɗin Kai na Sarrafa Girma ta Atomatik

COLA tana sake fasalin matsalar sarrafa girma ta atomatik a matsayin matsala mai ƙuntatawa ta ingantawa. Tana maye gurbin masu sarrafa girma masu zaman kansu da yawa da mai sarrafawa guda ɗaya, mai cibiyar sadarwa wanda ke da hangen nesa na duniya akan topology da aikin ƙananan ayyuka na aikace-aikacen.

4. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Za a iya wakilta babbar matsalar ingantawa a taƙaice kamar haka:

$$\min_{\{n_i\}} \sum_{i=1}^{M} C_i(n_i)$$ $$\text{ƙarƙashin sharadi: } L_{e2e}(\{n_i\}, \lambda) \leq L_{target}$$ $$n_i \in \mathbb{Z}^+$$ Ina:

• $M$: Adadin ƙananan ayyuka.
• $n_i$: Adadin raka'a albarkatu (misali, VMs) don ƙaramin sabis $i$.
• $C_i(n_i)$: Aikin farashi don ƙaramin sabis $i$ tare da raka'a $n_i$.
• $L_{e2e}$: Aikin jinkirin ƙarshe-zuwa-ƙarshe, ya dogara da duk $n_i$ da ƙarfin nauyin aiki $\lambda$.
• $L_{target}$: Manufar jinkirin SLO da ake so.

5. Sakamakon Gwaji & Kimantawa

An kimanta COLA sosai da wasu masu sarrafa girma ta atomatik na asali (na tushen amfani da na ML) akan Google Kubernetes Engine (GKE).

6. Tsarin Bincike & Misalin Hali

Hangen Nesa na Manazarcin: Ragewa ta Matakai Hudu

Fahimta ta Asali: Babban nasarar takardar ba sabon algorithm ne mai ban sha'awa ba, amma canji mai mahimmanci a cikin hangen nesa: ɗaukar dukan aikace-aikacen ƙananan ayyuka a matsayin tsarin guda ɗaya da za a inganta, ba tarin sassa masu zaman kansu ba. Wannan yayi kama da canjin da aka samu a hangen nesa na kwamfuta ta hanyar samfura kamar CycleGAN (Zhu et al., 2017), wanda ya wuce fassarar hoto biyu ta hanyar la'akari da daidaiton zagaye na duk yankin canji. COLA tana amfani da irin wannan "ka'idar daidaiton duniya" ga sarrafa albarkatu.

Kwararar Ma'ana: Hujjar tana da sauƙi mai jan hankali: 1) Mafi kyawun gida (daidaitawa kowane sabis) ya haɗa zuwa rashin inganci na duniya. 2) Saboda haka, yi amfani da manufa ta duniya (farashi) tare da ƙuntatawa ta duniya (jinkirin ƙarshe-zuwa-ƙarshe). 3) Tunda magance wannan akan layi yana da wuƙa sosai, magance shi ta hanyar bincike kashe layi kuma a turawa manufar. Kyawun yana cikin amfani da matsalar bandit don sanya binciken don mafi kyawun rarraba cikas ya zama mai inganci, dabarar da bincike mai yawa a cikin ƙarfafa koyo don ingantawar tsarin (misali, aiki daga RISELab na UC Berkeley) ke goyan bayan.

Ƙarfi & Aibobi: Ƙarfi: Sakamakon gwaji yana da kyau sosai—rage farashi na 19.3% adadi ne na matakin ɗakin zartarwa. Hanyar kashe layi tana da hankali, tana guje wa rashin kwanciyar hankali na lokacin aiki. Tsarin ba shi da alaƙa da dandamali. Aibobi: Achilles heel shine dogaro ga nauyin aiki na wakilci na kashe layi. A cikin aikace-aikace masu saurin ci gaba ko ƙarƙashin abubuwan zirga-zirgar "baƙar fata," manufar da aka riga aka lissafta na iya zama tsoho ko bala'i. Komawar takardar zuwa masu sarrafa girma ta atomatik na asali bandeji ne, ba magani ba, ga wannan matsalar ƙarfi. Bugu da ƙari, rikitarwar bincike mai yiwuwa tana da ƙarancin ma'auni tare da adadin ƙananan ayyuka, mai yuwuwa ta iyakance amfani da shi a cikin aikace-aikace masu girma sosai, masu rikitarwa.

Fahimta Mai Aiki: Ga masu gine-ginen girgije, saƙon yana bayyana: daina saita ƙofofin CPU da kansu. Saka hannun jari a cikin gina ko karɓar lura da aikin duniya da injin yanke shawara na tsakiya. Fara da hanyar haɗin gwiwa: yi amfani da falsafar COLA don ayyana mahimman sarƙoƙin sabis kuma a yi amfani da daidaitawar haɗin kai a can, yayin barin sabis masu zaman kansu marasa mahimmanci akan HPA na gargajiya. ROI, kamar yadda aka nuna, na iya zama sauri. Ya kamata masu bayar da sabis na girgije su lura; kayan aiki kamar GKE Autopilot suna buƙatar irin waɗannan yadudduka masu hankali na kaɗa don gaske cika alkawarin kayan aiki "masu sarrafawa".

7. Hangen Nesa na Aikace-aikace & Hanyoyin Gaba

Ka'idojin da ke bayan COLA suna da fa'ida mai faɗi fiye da daidaitawar VM na asali:

• Haɗin Kai na Albarkatu Da Yawa & Daidaitawa Daban-daban: Siffofin gaba na iya yanke shawara tare kan girman VM (ingantaccen ƙwaƙwalwar ajiya da na lissafi), rarraba GPU, har ma da sanyawa a cikin yankuna masu wadatarwa ko masu bayar da sabis na girgije don farashi da juriya.
• Haɗin Kai tare da Meshes na Sabis: Haɗa COLA tare da mesh na sabis (kamar Istio) zai samar da telemetry mai wadata (binciken matakin buƙatu, zane-zane na dogaro) har ma da ba da damar sarrafa kai tsaye kan hanyar zirga-zirga da karya da'ira a matsayin wani ɓangare na ingantawa.
• Daidaitawa akan Layi & Meta-Koyo: Babban iyakar bincike shine shawo kan iyakancewar kashe layi. Dabarun daga meta-koyo na iya ba da damar COLA ta daidaita manufarta akan layi da sauri bisa ra'ayi na ainihi, ko don bincika sabbin tsare-tsare cikin aminci a lokutan ƙarancin zirga-zirga.
• Manufofin Lissafi na Kore: Manufar ingantawa za a iya faɗaɗa ta don rage sawun carbon ko amfani da makamashi, daidaitawa da shirye-shiryen lissafi mai dorewa, ta hanyar haɗa bayanai daga tushe kamar aikin Sawun Carbon na Girgije.
• Kasuwar Manufofi: Don tsarin aikace-aikace na gama-gari (misali, kasuwanci ta e-commerce, watsa shirye-shiryen kafofin watsa labarai), za a iya raba ko sayar da manufofin COLA da aka riga aka inganta, yana rage buƙatar gudanar da horo na mutum ɗaya.

8. Nassoshi

1. Sachidananda, V., & Sivaraman, A. (2022). COLA: Collective Autoscaling for Cloud Microservices. arXiv preprint arXiv:2112.14845v3.
2. Zhu, J., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. Proceedings of the IEEE International Conference on Computer Vision (ICCV).
3. Burns, B., Grant, B., Oppenheimer, D., Brewer, E., & Wilkes, J. (2016). Borg, Omega, and Kubernetes. Queue, 14(1), 70–93.
4. Hoffman, M., Shahriari, B., & Aslanides, J. (2020). Addressing Function Approximation Error in Actor-Critic Methods. Proceedings of the 37th International Conference on Machine Learning (ICML). (Misalin RL na ci gaba mai dacewa don daidaitawa akan layi).
5. Cloud Carbon Footprint. (n.d.). An open source tool to measure and visualize the carbon footprint of cloud usage. Retrieved from https://www.cloudcarbonfootprint.org/.
6. Verma, A., Pedrosa, L., Korupolu, M., Oppenheimer, D., Tune, E., & Wilkes, J. (2015). Large-scale cluster management at Google with Borg. Proceedings of the European Conference on Computer Systems (EuroSys).