I-Telemetry ku-Junos ye-AI/ML Workloads
Umbhali: Shalini Mukherjee

Isingeniso

Njengoba ukuhamba kweqoqo le-AI kudinga amanethiwekhi angalahleki ane-throughput ephezulu kanye ne-latency ephansi, isici esibalulekile senethiwekhi ye-AI ukuqoqwa kwedatha yokuqapha. I-Junos Telemetry inika amandla ukuqapha kwe-granular yezinkomba zokusebenza ezibalulekile, okuhlanganisa imingcele kanye nezindawo zokubala zokulawula ukuminyana nokulinganisa ukulayisha kwe-traffic. Amaseshini e-gRPC asekela ukusakazwa kwedatha ye-telemetry. I-gRPC iwuhlaka lwesimanjemanje, lomthombo ovulekile, nokusebenza okuphezulu olwakhelwe kwezokuthutha ze-HTTP/2. Inika amandla amandla okusakaza omdabu aqondiswa kabili futhi ihlanganisa imethadatha yangokwezifiso eguquguqukayo kumaheda esicelo. Isinyathelo sokuqala ku-telemetry ukwazi ukuthi iyiphi idatha okufanele iqoqwe. Singabe sesihlaziya le datha ngamafomethi ahlukahlukene. Uma sesiqoqile imininingwane, kubalulekile ukuyethula ngendlela okulula ukuyibheka, ukuthatha izinqumo kanye nokwenza ngcono isevisi ehlinzekwayo. Kuleli phepha, sisebenzisa isitaki se-telemetry esihlanganisa i-Telegraf, i-InfluxDB, ne-Grafana. Lesi sitaki se-telemetry siqoqa idatha kusetshenziswa imodeli yokuphusha. Amamodeli okudonsa endabuko asebenzisa kakhulu izinsiza, adinga ukungenelela mathupha, futhi angafaka phakathi izikhala zolwazi kudatha ayiqoqayo. Amamodeli aphushayo anqoba le mikhawulo ngokuletha idatha ngokuvumelanayo. Bacebisa idatha ngokusebenzisa i-user-friendly tags kanye namagama. Uma idatha isikufomethi efundeka kakhudlwana, siyigcina kusizindalwazi futhi siyisebenzise ekubukeni okusebenzisanayo web isicelo sokuhlaziya inethiwekhi. Umfanekiso. 1 isibonisa ukuthi lesi sitaki sidizayinelwe kanjani ukuqoqwa kwedatha okusebenzayo, ukugcinwa, nokuboniswa, kusukela kumadivayisi enethiwekhi aphusha idatha kumqoqi kuya kudatha eboniswa kumadeshibhodi ukuze ihlaziywe.

Isitaki se-TIG

Sisebenzise iseva ye-Ubuntu ukufaka yonke isofthiwe kuhlanganise nesitaki se-TIG.

I-Telegraph
Ukuqoqa idatha, sisebenzisa i-Telegraf kuseva ye-Ubuntu esebenzisa i-22.04.2. Inguqulo ye-Telegraf esebenza kule demo ingu-1.28.5.
I-Telegraf iyi-ejenti yeseva eqhutshwa yi-plugin yokuqoqa nokubika amamethrikhi. Isebenzisa iprosesa plugins ukucebisa nokwenza idatha ibe yejwayelekile. Okukhiphayo plugins zisetshenziselwa ukuthumela le datha ezitolo zedatha ezahlukahlukene. Kulo mbhalo sisebenzisa ezimbili plugins: eyodwa eyezinzwa ze-openconfig kanti enye eyezinzwa zomdabu zaseJuniper.
I-InfluxDB
Ukugcina idatha kusizindalwazi sochungechunge lwesikhathi, sisebenzisa i-InfluxDB. I-plugin ephumayo ku-Telegraf ithumela idatha ku-InfluxDB, eyigcina ngendlela ephumelela kakhulu. Sisebenzisa i-V1.8 njengoba ingekho i-CLI ekhona ye-V2 nangaphezulu.
Grafana
I-Grafana isetshenziselwa ukubona le datha ngeso lengqondo. I-Grafana idonsa idatha ku-InfluxDB futhi ivumela abasebenzisi ukuthi bakhe amadeshibhodi anothile nasebenzisanayo. Lapha, sisebenzisa inguqulo 10.2.2.

Ukucushwa Ekushintsheni

Ukuze sisebenzise lesi sitaki, sidinga kuqala ukulungisa iswishi njengoba kuboniswe kuMfanekiso 2. Sisebenzise imbobo engu-50051. Noma iyiphi imbobo ingasetshenziswa lapha. Ngena ku-QFX switch bese wengeza ukucushwa okulandelayo.

Qaphela: Lokhu kulungiselelwa okwamalebhu/ama-POC njengoba igama eliyimfihlo lidluliselwa ngombhalo ocacile. Sebenzisa i-SSL ukuze ugweme lokhu.

Imvelo

Nginx
Lokhu kuyadingeka uma ungakwazi ukudalula itheku okungethwe kulo iGrafana. Isinyathelo esilandelayo ukufaka i-nginx kuseva ye-Ubuntu ukuze usebenze njenge-ejenti yommeleli ehlanekezelwe. Uma i-nginx isifakiwe, engeza imigqa ekhonjiswe kuMfanekiso 4 efayeleni “elizenzakalelayo” bese uhambisa ifayela ukusuka ku/etc/nginx kuya ku-/etc/nginx/sites-enabled.

Qinisekisa ukuthi i-firewall ilungisiwe ukuze inikeze ukufinyelela okugcwele kusevisi ye-nginx njengoba kukhonjisiwe kuMfanekiso 5.

Uma i-nginx isifakiwe futhi izinguquko ezidingekayo zenziwe, kufanele sikwazi ukufinyelela i-Grafana kusuka ku- web usebenzisa ikheli le-IP leseva ye-Ubuntu lapho wonke ama-software efakiwe.
Kukhona iphutha elincane ku-Grafana elingakuvumeli ukuthi usethe kabusha iphasiwedi ezenzakalelayo. Sebenzisa lezi zinyathelo uma uhlangabezana nalolu daba.
Izinyathelo okufanele zenziwe kuseva ye-Ubuntu ukusetha iphasiwedi ku-Grafana:

• Iya kokuthi /var/lib/grafana/grafana.db
• Faka i-sqllite3
  o sudo apt ukufaka sqlite3
• Sebenzisa lo myalo kutheminali yakho
  o sqlite3 grafana.db
•  Sqlite umyalo ngokushesha uyavula; sebenzisa lo mbuzo olandelayo:
  > susa kumsebenzisi lapho login = 'admin'
• Qala kabusha i-grafana bese uthayipha u-admin njengegama lomsebenzisi nephasiwedi. Icela iphasiwedi entsha.

Uma yonke i-software isifakiwe, dala ifayela le-config ku-Telegraf elizosiza ukudonsa idatha ye-telemetry ku-switch bese uyiphusha ku-InfluxDB.

I-Openconfig Sensor Plugin

Kuseva ye-Ubuntu, hlela ifayela /etc/telegraf/telegraf.conf ukwengeza konke okudingekayo. plugins kanye nezinzwa. Ngezinzwa ze-openconfig, sisebenzisa i-plugin ye-gNMI eboniswe kuMfanekiso 6. Ngezinjongo zedemo, engeza igama lokusingatha njengokuthi “spine1”, inombolo yembobo ethi “50051” esetshenziselwa i-gRPC, igama lomsebenzisi nephasiwedi yeswishi, kanye nenombolo. kwamasekhondi okuphinda kudayiwe uma yehluleka.
Esigabeni sokubhaliselwe, engeza igama eliyingqayizivele, "cpu" yale nzwa ethile, indlela yenzwa, kanye nesikhawu sesikhathi sokuthatha le datha kuswishi. Engeza i-plugin inputs.gnmi efanayo kanye ne-input.gnmi.subscription yazo zonke izinzwa ezivuliwe zokuhlanganisa. (Umfanekiso 6)

I-plugin yenzwa yomdabu

Lena i-plugin ye-Juniper telemetry interface esetshenziselwa izinzwa zomdabu. Kufayela elifanayo le-telegraf.conf, engeza okokufaka kwe-plugin yenzwa yomdabu.jti_openconfig_telemetry lapho izinkambu zicishe zifane ne-openconfig. Sebenzisa i-ID yeklayenti ehlukile kuyo yonke inzwa; lapha, sisebenzisa "telegraf3". Igama eliyingqayizivele elisetshenziswe lapha kule nzwa lithi “mem” (Umfanekiso 7).

Okokugcina, engeza i-plugin outputs.influxdb ukuze uthumele le datha yenzwa ku-InfluxDB. Lapha, isizindalwazi siqanjwe ngokuthi “telegraf” negama lomsebenzisi elithi “influx” kanye nephasiwedi “influxdb” (Umfanekiso 8).

Uma usulihlelile ifayela le-telegraf.conf, qala kabusha isevisi ye-telegraf. Manje, hlola i-InfluxDB CLI ukuze uqiniseke ukuthi izilinganiso zenzelwe zonke izinzwa ezihlukile. Thayipha okuthi “influx” ukuze ufake i-InfluxDB CLI.

Njengoba kubonakala kuMfanekiso. 9, faka umyalo we-influxDB bese usebenzisa i-database ethi “telegraf”. Wonke amagama ahlukile anikezwe izinzwa abhalwe njengezilinganiso.
Ukuze ubone ukuphuma kwanoma yisiphi isilinganiso esisodwa, ukuze uqiniseke ukuthi ifayela le-telegraf lilungile futhi inzwa iyasebenza, sebenzisa umyalo othi "khetha * kumkhawulo we-cpu 1" njengoba kukhonjisiwe kuMfanekiso 10.

Ngaso sonke isikhathi uma kwenziwa izinguquko efayeleni le-telegraf.conf, qiniseka ukuthi umisa i-InfluxDB, qala kabusha i-Telegraf, bese uqala i-InfluxDB.
Ngena ku-Grafana kusuka kusiphequluli bese udala amadeshibhodi ngemva kokuqinisekisa ukuthi idatha iqoqwa ngendlela efanele.
Iya kokuthi Ukuxhumana > I-InfuxDB > Engeza umthombo wedatha omusha.

1. Nikeza igama kulo mthombo wedatha. Kule demo ithi “test-1”.
2.  Ngaphansi kwesigaba se-HTTP, sebenzisa iseva ye-Ubuntu IP kanye nembobo ye-8086.
   
3. Emininingwaneni ye-InfluxDB, sebenzisa igama lesizindalwazi elifanayo, “telegraf,” bese unikeza igama lomsebenzisi nephasiwedi yeseva ye-Ubuntu.
4. Chofoza okuthi Londoloza futhi uhlole. Qinisekisa ukuthi uyawubona umlayezo othi, “kuphumelele”.
   
5. Uma umthombo wedatha usungezwe ngempumelelo, hamba kokuthi Amadeshibhodi bese uchofoza Okusha. Masidale amadeshibhodi ambalwa abalulekile kumthwalo we-AI/ML kumodi yomhleli.

ExampLes Of Amagrafu enzwa

Okulandelayo yi-exampokuncane kwezinye izinto zokubala ezinkulu ezibalulekile ekuqapheni inethiwekhi ye-AI/ML.
Iphesentitagukusetshenziswa kwe-interface ye-ingress et-0/0/0 kumgogodla-1

• Khetha umthombo wedatha njengokuhlola-1.
• Esigabeni esithi FROM, khetha ukulinganisa njengokuthi “interface”. Leli igama eliyingqayizivele elisetshenziselwa le ndlela yenzwa.
• Esigabeni esithi LAPHO, khetha idivayisi::tag, futhi ku tag value, khetha igama lomethuleli weswishi, okungukuthi, umgogodla1.
• Esigabeni esithi KHETHA, khetha igatsha lenzwa ofuna ukuliqapha; kulesi simo khetha “inkambu(/interfaces/interface[if_name='et-0/0/0']/state/counters/if_in_1s_octets)”. Manje esigabeni esifanayo, chofoza ku-“+” bese wengeza lesi sibalo sokubala (/50000000000 * 100). Sibala iphesentitage ukusetshenziswa kwe-interface ye-400G.
• Qiniseka ukuthi i-FORMAT "iwuchungechunge lwesikhathi," bese uqamba igrafu esigabeni se-ALIAS.

Isikhathi esiphezulu sebhafa kunoma yimuphi ulayini

• Khetha umthombo wedatha njengokuhlola-1.
• Esigabeni esithi FROM, khetha ukulinganisa njengokuthi “buffer.”
• Esigabeni esithi LAPHO, kunezinkambu ezintathu okufanele zigcwaliswe. Khetha idivayisi::tag, futhi ku tag inani khetha igama lomethuleli weswishi (okungukuthi umgogodla-1); FUTHI ukhethe /cos/interfaces/interface/@name::tag bese ukhetha isixhumi esibonakalayo (okungukuthi et- 0/0/0); FUTHI ukhethe ulayini futhi, /cos/interfaces/interface/queues/queue/@queue::tag bese ukhetha inombolo yomugqa 4.
• Esigabeni esithi KHETHA, khetha igatsha lenzwa ofuna ukuliqapha; kulesi simo khetha “inkambu(/cos/interfaces/interface/queues/queue/PeakBufferOccupancy).”
• Qinisekisa ukuthi i-FORMAT "iwuchungechunge lwesikhathi" futhi uqambe igrafu esigabeni se-ALIAS.

Ungakwazi ukuhlanganisa idatha yezindawo zokusebenzelana eziningi kugrafu efanayo njengoba ibonakala kuMfanekiso 17 ka-et-0/0/0, et-0/0/1, et-0/0/2 njll.

I-PFC ne-ECN isho okuphuma kokunye

Ukuze uthole okuphuma kokuphuma kuyo okumaphakathi (umehluko enani phakathi kwebanga lesikhathi), sebenzisa imodi yombuzo ongahluziwe.
Lona umbuzo we-influx esiwusebenzisile ukuze sithole okuphuma kokunye okumaphakathi phakathi kwamanani amabili e-PFC ku-et-0/0/0 ye-Spine-1 isekhondi.
KHETHA okuphuma kokunye(isho(“/interfaces/interface[if_name='et-0/0/0′]/state/pfc-counter/tx_pkts”), 1s) KUSUKA ““interface” LAPHO (“idivayisi”::tag = 'Umgogodla-1') KANYE neQEMBU le-$timeFilter NGESIKHATHI($interval)

KHETHA okuphuma kuyo (isho(“/interfaces/interface[if_name='et-0/0/8′]/state/error-counters/ecn_ce_marked_pkts”), 1s) KUSUKA ““interface” LAPHO (“idivayisi”::tag = 'Umgogodla-1') KANYE neQEMBU le-$timeFilter NGESIKHATHI($interval)

Amaphutha wensiza yokufaka asho okuphuma kokunye

Umbuzo ongahluziwe wamaphutha esisetshenziswa kusho okuphuma kokunye:
KHETHA okuphuma ku-derivative(isho(“/interfaces/interface[if_name='et-0/0/0′]/state/error-counters/if_in_resource_errors”), 1s) KUSUKA ““interface” LAPHO (“idivayisi”::tag = 'Umgogodla-1') KANYE neQEMBU le-$timeFilter NGESIKHATHI($interval)

Amaconsi omsila asho okuphuma kokunye

Umbuzo ongahluziwe wokuphuma komsila usho ukuthi:
KHETHA okuphuma kokunye(mean(“/cos/interfaces/interface/queue/queue/tailDropBytes”), 1s) KUSUKA “ku-bu” LAPHO (“idivayisi”::tag = 'Leaf-1' KANYE “/cos/interfaces/interface/@name”::tag = 'et-0/0/0' KANYE “/cos/interfaces/interface/queues/queue/@queue”::tag = '4') KANYE neQEMBU le-$timeFilter NGESIKHATHI($__interval) gcwalisa(null)
 Ukusetshenziswa kwe-CPU

• Khetha umthombo wedatha njengokuhlola-1.
• Esigabeni esithi FROM, khetha ukulinganisa njengokuthi “newcpu”
• KU-LAPHO, kunemikhakha emithathu okufanele igcwaliswe. Khetha idivayisi::tag futhi kwe tag inani khetha igama lomethuleli weswishi (okungukuthi umgogodla-1). FUTHI kokuthi /izingxenye/ingxenye/izakhiwo/izakhiwo/igama:tag, bese ukhetha i-cpuutilization-total KANYE egameni::tag khetha RE0.
• Esigabeni esithi KHETHA, khetha igatsha lenzwa ofuna ukuliqapha. Kulokhu, khetha "inkambu(isimo/inani)".

Umbuzo ongahluziwe wokuthola okuphuma kokunye okungenanegethivu komsila kuyehla ekushintsheni okuningi kuzixhumanisi eziningi ngamabhithi/isekhondi.
KHETHA non_negative_derivative(mean(“/cos/interfaces/interface/queue/queue/tailDropBytes”), 1s)*8 KUSUKA “BU” LAPHO (idivayisi::tag =~ /^Spine-[1-2]$/) kanye (“/cos/interfaces/interface/@name”::tag =~ /et-0\/0\/[0-9]/ noma “/cos/interfaces/interface/@name”::tag=~/et-0\/0\/1[0-5]/) KANYE ne-$timeFilter GROUP BY time($__interval),idivayisi::tag gcwalisa(null)

Lawa kwakungamanye ama-exampokuncane kwamagrafu angadalelwa ukuqapha inethiwekhi ye-AI/ML.

Isifinyezo

Leli phepha libonisa indlela yokudonsa idatha ye-telemetry nokuyibona ngeso lengqondo ngokwakha amagrafu. Leli phepha likhuluma ngokuqondile ngezinzwa ze-AI/ML, zomdabu kanye ne-openconfig kodwa ukusetha kungasetshenziswa kuzo zonke izinhlobo zezinzwa. Futhi sifake izixazululo zezinkinga eziningi ongase ubhekane nazo ngenkathi udala ukusethwa. Izinyathelo nemiphumela evezwe kuleli phepha icaciselwe izinguqulo zesitaki se-TIG okukhulunywe ngazo ekuqaleni. Ingashintsha kuye ngenguqulo yesofthiwe, izinzwa kanye nenguqulo ye-Junos.

Izithenjwa

I-Juniper Yang Data Model Explorer yazo zonke izinketho zezinzwa
https://apps.juniper.net/ydm-explorer/
Iforamu ye-Openconfig yezinzwa ze-openconfig
https://www.openconfig.net/projects/models/

Ikomkhulu Lebhizinisi Nezokuthengisa
Inkampani Juniper Networks, Inc.
1133 Innovation Way
Sunnyvale, CA 94089 USA
Ucingo: 888. JUNIPER (888.586.4737)
noma +1.408.745.2000
Ifeksi: +1.408.745.2100
www.juniper.net
Ikomkhulu le-APAC ne-EMEA
IJuniper Networks International BV
I-Boeing Avenue 240
1119 PZ Schiphol-Rijk
E-Amsterdam, eNetherlands
Ucingo: +31.207.125.700
Ifeksi: +31.207.125.701
Copyright 2023 Juniper Networks. Inc. Amalungelo e-Ail agodliwe. IJuniper Networks, ilogo yeJuniper Networks, iJuniper, iJunos, nezinye izimpawu zokuthengisa yizimpawu zokuthengisa ezibhalisiwe zeJuniper Networks. inc. kanye/noma izinhlaka zayo e-United States nakwamanye amazwe. Amanye amagama angaba yizimpawu zokuthengisa zabanikazi bazo. IJuniper Networks ayinaso isibopho sanoma yikuphi ukungalungi kulo mbhalo. IJuniper Networks iyaligodla ilungelo lokushintsha. guqula. dlulisa, noma ubuyekeze lokhu kushicilelwa ngaphandle kwesaziso.
Thumela impendulo ku: design-center-comments@juniper.net V1.0/240807/ejm5-telemetry-junos-ai-ml

Amadokhumenti / Izinsiza

I-Juniper NETWORKS Telemetry In Junos ye-AI ML Workloads Software [pdf] Umhlahlandlela Womsebenzisi I-Telemetry In Junos ye-AI ML Workloads Software, i-Junos ye-AI ML Workloads Software, i-AI ML Workloads Software, i-Workloads Software, isofthiwe

Izithenjwa

• Imaniwali yosebenzisayo