Intel LogoIntel ® Ethernet 700 Series
Jagorar Tunawa Ayyukan Linux
NEX Cloud Networking Group (NCNG)
Littafin 1.2
Disamba 2024

Abubuwan da ke ciki boye
1 Tarihin Bita
2 Gabatarwa
3 Jerin abubuwan dubawa na farko
4 Ma'aunin Aiki na Tushe da Hanyar Tunatarwa
5 Tuning i40e Driver Saituna
6 Tuning Platform (I40e Ba Takamaiman ba)
7 Adafta Bonding
8 Magance Matsalar Aiki
9 Shawarwari don Yanayin Ayyuka na gama-gari
10 Takardu / Albarkatu
10.1 Magana

Tarihin Bita

Bita Kwanan wata Sharhi
1.2 Disamba 2024 · Ƙara ƙarin Jagorar Gudanar da Wuta.
An ƙara Intel * Turbo Boost.
· Ƙara bayanan baya na na'urar sadarwa.
· Ƙara Ƙaddamar Takamaiman Kanfigareshan da Tunatarwa.
An Ƙara 4th Generation Intel* %eon* Masu Sarrafa Ma'auni.
· An ƙara AMD EPYC.
· Sabunta Duba Ƙarfin Hardware System.
· An sabunta iPerf2.
· An sabunta iPerf3.
Ɗaukaka Tx/Rx Queues.
· Sabunta Tsagewar Tattalin Arziki.
Girman Zoben da aka sabunta.
Tuning Platform da aka sabunta (I40e Ba Takamaiman ba).
· Sabunta saitunan BIOS.
· An sabunta Ikon Jihar C.
· Sabunta Ma'auni na Mitar CPU.
· Sabunta Saitunan Aikace-aikace.
· Sabunta Tsarin Aiki/Saitunan Kernel.
· Ƙaddamar da IP da aka sabunta.
· An sabunta Low Latency.
Agusta 2023 Canje-canje ga wannan takarda sun haɗa da:
· Ƙara Abubuwan da ke da alaƙa.
An Ƙara Tabbatar Kunshin DDP yana Loda Da kyau.
An ƙara iPerf2.
An ƙara iPerf3.
· Ƙara netperf.
· Sabunta IRQ Affinity.
· Ƙara Tx/Rx Queues.
Girman Zoben da aka sabunta.
· Ƙara Firam ɗin Jumbo.
Ƙaddamar da Adafta Bonding.
· An ƙara Intel svr-info Tool.
1.0 Maris 2016 Sakin Farko (Intel Jama'a).

Gabatarwa

An yi nufin wannan jagorar don samar da jagora don daidaita yanayin don ingantaccen aikin sadarwar ta amfani da Intel ® Ethernet 700 Series NICs a cikin mahallin Linux. Yana maida hankali kan hardware, direba, da yanayin tsarin aiki da saituna waɗanda zasu inganta aikin cibiyar sadarwa. Ya kamata a lura cewa aikin sadarwar zai iya shafar kowane adadin tasirin waje, kawai mafi yawan abubuwan da aka fi sani da ban mamaki a cikin wannan jagorar.
1.1 Abubuwan da ke da alaƙa

Jerin abubuwan dubawa na farko

2.1 Sabunta Sabbin Direba/ Firmware
Bincika nau'ikan direba/firmware ta amfani da ethtool -i ethx.
Sabunta mai zuwa kamar yadda ake buƙata:

2.2 Karanta KARANTA
Bincika abubuwan da aka sani kuma sami sabbin umarnin daidaitawa daga README file kunshe a cikin i40e tushen kunshin.
2.3 Bincika cewa Ramin PCI Express (PCIe) ɗinku x8 ne
Wasu ramukan PCIe x8 an daidaita su azaman ramukan x4. Waɗannan ramummuka suna da ƙarancin bandwidth don cikakken ƙimar layi tare da tashar jiragen ruwa biyu da na'urorin tashar tashar quad. Bugu da kari, idan kun sanya adaftar mai iya PCIe v3.0 a cikin ramin PCIe v2.x, ba za ku iya samun cikakken bandwidth ba. Direban na'urar software ya gano wannan yanayin kuma ya rubuta saƙo mai zuwa a cikin log ɗin tsarin:
PCI-Express bandwidth da ke akwai don wannan katin bai isa ba don kyakkyawan aiki. Don ingantaccen aiki ana buƙatar ramin x8 PCI-Express.
Idan wannan kuskuren ya faru, matsar da adaftar ku zuwa ramin PCIe v3.0 x8 na gaskiya don warware matsalar.
2.4 Duba Ƙarfin Hardware na System
A 10 Gbps, 25 Gbps, da 40 Gbps Ethernet, akwai wasu ƙananan CPU da buƙatun tsarin. Gabaɗaya, na'ura mai sarrafa nau'in uwar garken zamani da ingantaccen tsarin ƙwaƙwalwar ajiya don dandamali ya kamata ya isa, amma buƙatun sun bambanta dangane da nauyin aikinku. Duk tashoshi na ƙwaƙwalwar ajiya yakamata a cika jama'a kuma yanayin aikin ƙwaƙwalwar ajiya yakamata a kunna a cikin BIOS. Tabbatar cewa CPU ɗinku da daidaitawar ƙwaƙwalwar ajiya suna da ikon tallafawa matakin aikin hanyar sadarwa da kuke buƙata don aikinku.
NOTE
XL710 shine mai sarrafa 40 GbE. Adaftar 2 x 40 GbE da ke amfani da wannan mai sarrafa ba a yi niyya ta zama 2 x 40 GbE ba amma 1 x 40 GbE tare da tashar jiragen ruwa mai aiki. Lokacin ƙoƙarin yin amfani da zirga-zirgar layin layi wanda ya haɗa da tashoshin jiragen ruwa biyu, canjin na ciki ya cika kuma haɗin haɗin kai tsakanin tashoshin biyu yana iyakance ga jimillar SO Gbps.
2.4.1 Kernel Boot Parameters
Idan Intel® Virtualization Technology for Directed I/O (Intel® VT-d) aka kunna a cikin BIOS, Intel yana ba da shawarar cewa IOMMU ta kasance cikin yanayin wucewa don ingantaccen aikin cibiyar sadarwa. Wannan yana kawar da sama da DMA akan zirga-zirgar zirga-zirga yayin da yake ba da damar Injinan Virtual (VMs) don har yanzu suna da fa'idodin Intel® VT-d. Ana yin wannan ta ƙara layin da ke gaba zuwa sigogin taya na kernel: fommu-pt.
2.5 Tabbatar Kunshin DDP yana Loda Da kyau
140ea da 140eb direbobin tushe ba su da tallafi kai tsaye don Keɓancewar Na'urar Dynamic (DDP). Don amfani da DDP tare da na'urori 700 Series, DDP profile za a iya amfani da aikace-aikacen testpmd.
Don cikakkun bayanai akan DDP profiles, da kuma yadda ake amfani da DDP profile tare da testpmd akan na'urori 700 Series, koma zuwa Intel® Ethernet 700 Series Dynamic Device Personalization (DDP) Jagoran fasaha.
Don tabbatar da ko DDP profile an yi lodi cikin nasara:
testpmd> ddp samun lissafin 0 Profile lamba shine: 1
NOTE
Idan profile lamba shine 0, ba a ɗora fakitin DDP ba. A cikin lamarin kuskuren lodin fakitin DDP, na'urar ta yi kasala zuwa yanayin aminci kuma ba a samun fasalulluka da yawa. Idan akwai kurakurai masu alaƙa da lodin kunshin DDP, zai haifar da al'amuran aiki. Don matakan magance matsala, koma zuwa Inte/* Ethernet 700 Series Dynamic Device Personalization (DDP) Jagoran fasaha.

Ma'aunin Aiki na Tushe da Hanyar Tunatarwa

3.1 Alamar Ayyukan Sadarwar Sadarwa
Kafin fara motsa jiki na daidaitawa, yana da mahimmanci a sami kyakkyawan ma'aunin tushe na aikin cibiyar sadarwar ku. Yawancin lokaci ban da samun ma'aunin farko na takamaiman aikace-aikacenku/nauyin aikinku, yana da kyau a yi amfani da ma'aunin aikin cibiyar sadarwa don tabbatar da cewa na'urar sadarwar ku tana cikin kyakkyawan yanayi.
Don haɓaka tsarin guda ɗaya, netperf ko iperf da NetPIPE duk ingantaccen kayan aikin buɗe tushen kyauta ne waɗanda ke ba ku damar ƙarfafa haɗin gwiwa da gano matsalolin aiki.
Netperf yana da ƙarfi don duka kayan aiki da gwajin latency. NetPIPE takamaiman kayan aiki ne na latency amma ana iya haɗa shi don kowane irin yanayi.
NOTE
Gwajin TCP_RR a cikin netperf yana dawo da latency a cikin ƙimar ma'amaloli/sec. Wannan lambar zagayawa ce. Ana iya ƙididdige latency ta hanya ɗaya ta amfani da ma'auni mai zuwa:
Latency(amfani) = (1⁄2) / [Ma'amaloli/sek] * 1,000,000
3.1.1 iPerf2
Intel yana ba da shawarar iperf2 akan iperf3 don mafi yawan yanayin ƙima saboda sauƙin amfani da goyan bayan zaren da yawa a cikin misalin aikace-aikacen guda ɗaya. Intel yana ba da shawarar yin aiki tare da zaɓi na -P tare da zaren 2-4 don haɗin haɗin 25G da kewayen zaren 4-6 don haɗin 40G.

  • Don gudanar da zirga-zirgar hanya ɗaya daga abokin ciniki zuwa uwar garken: Umurnin uwar garken exampda: arf2 -s
    Umurnin abokin ciniki exampda: tsirf2 -c -P
  • Don gudanar da zirga-zirgar ababen hawa biyu daga abokin ciniki zuwa uwar garken (kuma akasin haka): Umurnin uwar garken exampda: iperf2 –s –p
    Umurnin abokin ciniki exampda:
    aiki 2 -c -p -P --cikakken-duplex KO
    aiki 2 -c -p -P –d

NOTE
Duk zaɓuɓɓukan -cikakken-duplex da -d a cikin iperf2 suna ba mai amfani damar yin gwaji na biyu. Koyaya, zaɓi - cikakken-duplex musamman yana mai da hankali kan cikakken gwajin duplex.
NOTE
Lokacin gwada iperf2 a cikin tashoshin uwar garken da yawa, za a iya ƙara tutar -d zuwa umarnin uwar garken don gudanar da duk zaman sabar a bango daga tagar tasha ɗaya. Hakanan za'a iya amfani da tutar -d lokacin da aka saka umarnin uwar garke a cikin madauki a cikin rubutun.
NOTE
Lokacin gudanar da gwajin kayan aikin cibiyar sadarwa tare da rafi / zaren guda ɗaya (misaliample: P1), na'urori na AMD bazai samar da kayan aiki da ake tsammani ba, musamman maɗaukakin bandwidth NICs (idan gudun shine> = bandwidth 25G). Sakamakon haka, ana buƙatar haɗa aikace-aikacen zuwa takamaiman maƙallan don cimma babban abin da ake samu. Duba Saitunan Aikace-aikace a shafi na 22.
3.1.2 iPerf3
Idan aka yi amfani da iperf3, ana buƙatar lokuta da yawa na aikace-aikacen don ɗaukar advantage na zaren da yawa, RSS, da layukan hardware. Intel yana ba da shawarar yin aiki tare da lokutan aikace-aikacen 2-4 don haɗin haɗin 25G da kuma kusan zaman 4-6 don haɗin 40G. Kowane zama yakamata ya ƙayyade ƙimar tashar tashar TCP ta musamman ta amfani da zaɓi -p.

  • Don gudanar da zirga-zirgar hanya ɗaya daga abokin ciniki zuwa uwar garken:
    Umurnin uwar garken misaliampda:
    kirarf3 -s -p
    Umurnin abokin ciniki exampda:
    kirarf3 -c -p
  • Don gudanar da zirga-zirgar ababen hawa biyu daga abokin ciniki zuwa uwar garken (kuma akasin haka):
    Umurnin uwar garken misaliampda:
    kirarf3 –s –p
    Umurnin abokin ciniki exampda: tsirf3 -c -p -P --bidir
  • Don fara lokuta da yawa (zaren) na iperf3, shawarar ita ce a yi amfani da madauki don taswirar zaren zuwa tashoshin TCP da gudanar da iperf3 a bango ta amfani da & don ƙirƙirar matakai da yawa a layi daya.
    Umurnin uwar garken misaliample, fara 4 zaren: tashar jiragen ruwa =”; don ni a cikin {0..3}; yi tashar jiragen ruwa=520$i; bash -c "iperf3 -s -p $ tashar jiragen ruwa &"; yi; Umurnin abokin ciniki example, fara 4 zaren - Canja wurin tashar gwaji =""; don ni a cikin {0..3}; yi tashar jiragen ruwa=520$i; bash -c "iperf3 -c $ serverIP -p $ tashar jiragen ruwa &"; yi; Umurnin abokin ciniki example, fara 4 zaren – Karɓi tashar gwaji =””; don ni a cikin {0..3}; yi tashar jiragen ruwa=520$i; bash -c "iperf3 -R -c $serverIP -p $ tashar jiragen ruwa &"; yi; Don haɗin 40G, ƙara madauki don ƙirƙirar har zuwa lokuta 6 / zaren.

NOTE
Lokacin gudanar da gwajin kayan aikin cibiyar sadarwa tare da rafi / zaren guda ɗaya (misaliample: P1), AMD na'urori masu sarrafawa bazai samar da kayan aiki da ake tsammani ba, musamman mafi girma bandwidth
NICs (idan gudun shine> = bandwidth 25G). Sakamakon haka, ana buƙatar haɗa aikace-aikacen zuwa takamaiman maƙallan don cimma babban abin da ake samu. Duba Saitunan Aikace-aikace a shafi na 22 da AMD EPYC a shafi na 26.
3.1.3 netperf
Kayan aikin netperf shine zaɓi mai ƙarfi don duka kayan aiki da gwajin latency.

  • Gwajin TCP_STREAM a cikin netperf yana auna iyawar kayan aiki na na'urar. Umurnin uwar garken misaliample: netserver Client umurnin exampda: netperf -t TCP_STREAM -l 30 -H
  • Gwajin TCP_RR a cikin netperf yana dawo da latency a cikin ƙimar ma'amaloli/daƙiƙa. Wannan lambar zagayawa ce. Ana ba da shawarar yin amfani da zaɓin -T x,x, kasance x na gida ne na na'urar. Ana iya ƙididdige latency ta hanya ɗaya ta amfani da: Latency(usec)=(1⁄2)/ [Ma'amaloli/sec]*1,000,\Umarnin uwar garke exampda: netserver
    Umurnin abokin ciniki exampda: netperf -t TCP_RR -l 30 -H - da x, x
  • Don fara lokuta da yawa (zaren) na netperf, shawarar ita ce a yi amfani da madauki don taswirar zaren zuwa tashar jiragen ruwa na TCP da gudanar da netperf a bango ta amfani da & don ƙirƙirar matakai da yawa a layi daya.
    Umurnin uwar garken misaliample, fara zaren 8:
    tashar jiragen ruwa ="; don ni a cikin {0..7}; yi tashar jiragen ruwa=520$i; bash -c "netserver -L $ uwar garkenIP -p $ tashar jiragen ruwa &"; yi;
    Umurnin abokin ciniki example, fara 8 zaren: tashar jiragen ruwa =”; don ni a cikin {0..7}; yi tashar jiragen ruwa=520$i; bash -c "netperf -H $ uwar garken IP -p $ tashar jiragen ruwa -t TCP_STREAM -l 30 &"; yi;

3.2 Hanyar Tunatarwa
Mayar da hankali ga sauyi guda ɗaya a lokaci guda don ku san irin tasirin kowane canji ke haifarwa ga gwajin ku. Da karin dabarar da kuke cikin tsarin daidaitawa, da sauƙin zai kasance don ganowa da magance abubuwan da ke haifar da cikas.

Tuning i40e Driver Saituna

4.1 Alamar IRQ
Haɓaka alaƙar IRQ don katse hanyoyin layin sadarwa daban-daban ana haɗa su zuwa nau'ikan CPU daban-daban na iya yin babban tasiri akan aiki, musamman gwaje-gwajen kayan aikin multithread.
Don daidaita alaƙar IRQ, dakatar da irqbalance sannan ko dai yi amfani da rubutun set_irq_affinity daga fakitin tushen i40e ko layin fil da hannu. Kashe ma'auni na IRQ-sarari mai amfani don kunna layin layi:

  • systemctl kashe irqbalance
  • systemctl dakatar da irqbalance
    Amfani da rubutun set_irq_affinity daga fakitin tushen i40e (an shawarta):
  • Don amfani da duk abin da ake bukata:
    [hanyar-zuwa-i40epackage]/scripts/set_irq_affinity -X duk ethX
  • Don amfani da kawai cores akan soket na NUMA na gida: [hanyar-zuwa-i40epackage]/scripts/set_irq_affinity -X local ethX
  • Hakanan zaka iya zaɓar kewayon muryoyi. Ka guji amfani da cpu0 saboda yana gudanar da ayyukan ƙidayar lokaci. [hanyar-zuwa-i40epackage]/scripts/set_irq_affinity 1-2 ethX

NOTE
Rubutun alaƙa yana ba da damar Gudanar da Fakitin Watsawa (XPS) azaman ɓangare na aiwatar da pinning lokacin da zaɓi -x ya ƙayyade. Lokacin da aka kunna XPS, Intel yana ba da shawarar ka kashe irqbalance, saboda ma'aunin kwaya tare da XPS na iya haifar da aiki maras tabbas. Rubutun alaƙa yana kashe XPS lokacin da zaɓi -X ya ƙayyade. Kashe XPS da ba da damar layukan ƙima suna da fa'ida ga nauyin aiki inda aka sami mafi kyawun aiki lokacin da aka sami sabis na zirga-zirgar Tx da Rx akan layi (s) guda biyu.
Haɓaka layukan ma'auni a cikin Linux ya haɗa da daidaita sigogin direba na cibiyar sadarwa don ba da damar karɓar layukan ƙima (Rx) da jerin gwano (Tx) don masu adaftar cibiyar sadarwa masu tallafi.
NOTE

  • Layukan simmetric siffa ce ta hanyar sadarwa ta ci gaba, kuma ba duka jerin adaftan cibiyar sadarwa 700 ko direbobi ke goyan bayansu ba.
  • Tabbatar cewa kana da direba da goyon bayan hardware kafin yunƙurin daidaita jerin gwano.

Don saita jerin gwano, bi waɗannan matakan gabaɗaya:

  1. Shirya Kanfigareshan Interface File: Yi amfani da editan rubutu (misaliample, vi, nano, ko gedit) don shirya saitin mu'amalar hanyar sadarwa file. The file yawanci yana ƙarƙashin /etc/sysconfig/network-scripts/ directory kuma yana da suna kamar ifcfg-ethX, inda ethX shine sunan cibiyar sadarwar ku.
  2. Ƙara Ma'auni na Simmetric Queue. Ƙara layukan da ke biyowa zuwa saitin mu'amalar cibiyar sadarwa file: ETHTOOL_OPTS=”rx-line-queues 8 tx-queues 8″
  3. Sake kunna sabis na hanyar sadarwa.
    Bayan yin canje-canje, sake kunna sabis na cibiyar sadarwa don amfani da sabon saitin. sudo systemctl sake kunna cibiyar sadarwa

Da hannu:

  • Nemo masu sarrafawa da ke haɗe zuwa kowane kumburi ta amfani da: numactl –hardware lscpu
  • Nemo abin rufe fuska ga kowane na'urori masu sarrafawa:
  • Tsammanin ƙira 0-11 don kumburi 0: [1,2,4,8,10,20,40,80,100,200,400,800]
  • Nemo IRQs ɗin da aka sanya wa tashar jiragen ruwa da ake sanyawa: grep ethX /proc/katsewa kuma lura da ƙimar IRQ Don ex.ample, 181-192 don vectors 12 da aka ɗora.
  • Ƙimar ƙimar alaƙar SMP a cikin shigarwar IRQ daidai. Lura cewa ana buƙatar yin wannan don kowane shigarwar IRQ: echo 1> /proc/irq/181/smp_affinity echo 2> /proc/irq/182/smp_affinity echo 4> /proc/irq/183/smp_affinity Show IRQ affinity:
  • Don nuna alaƙar IRQ ga duk mahimman bayanai: /scripts/set_irq_affinity -s ethX
  • Don nuna maɓalli kawai akan soket na NUMA na gida: /scripts/set_irq_affinity -s local ethX
  • Hakanan zaka iya zaɓar kewayon muryoyi: /scripts/set_irq_affinity -s 40-0-8,16 ethX

NOTE
Rubutun saitin_irq_affinity yana goyan bayan tutar -s a sigar direban i40e 2.16.11 da kuma daga baya.
4.2 Tx/Rx Queues
Tsohuwar adadin layukan da aka kunna don kowane tashar Ethernet ta direba a farkon farawa daidai yake da jimillar adadin CPUs da ake samu a dandalin. Wannan yana aiki da kyau don dandamali da yawa da daidaitawar aikin aiki. Koyaya, a cikin dandamali masu ƙididdige ƙididdiga masu girma da / ko babban tashar tashar Ethernet, wannan daidaitawar na iya haifar da takaddamar albarkatu. Don haka, yana iya zama dole a wasu lokuta don gyara tsohowar kowane tashar jiragen ruwa a cikin tsarin.
Adadin tsoho na layin Tx/Rx na iya bambanta dangane da takamaiman samfurin da sigar direba. Ana iya daidaita adadin layukan ta amfani da umarnin ethtool -L da aka jera a ƙasa.
NOTE
A cikin waɗannan lokuta, Intel yana ba da shawarar cewa ku rage adadin tsoffin layi na kowane tashar jiragen ruwa zuwa fiye da adadin CPUs da ke akwai a cikin NUMA node na gida zuwa tashar adaftar. A wasu lokuta, lokacin ƙoƙarin daidaita albarkatu a kan manyan ayyukan ƙidayar tashar jiragen ruwa, yana iya zama dole a rage wannan lambar har ma da gaba.
Don gyara tsarin layi:
Mai zuwa example saita tashar jiragen ruwa zuwa 32 Tx/Rx jerin gwano: ethtool -L ethX hade 32
Exampda fitarwa:
ethtool -l ethX
Simitocin tashoshi don ethX: An riga an saita iyakar:
Saukewa: 96
Saukewa: 96
Sauran: 1
Hade: 96
Saitunan kayan masarufi na yanzu:
Saukewa: 0
Saukewa: 0
Sauran: 1
Hade: 32
4.3 Tsage Tsaye
Matsakaicin katsewa mai daidaitawa yana kunne ta tsohuwa, kuma an ƙirƙira shi don samar da madaidaicin hanya tsakanin ƙarancin amfani da CPU da babban aiki. Koyaya, zaku iya gwada kunna saitunan katsewa da hannu don dacewa da yanayin amfanin ku.
Matsakaicin 0-235 micro seconds yana ba da ingantaccen kewayon katsewa 4,310 zuwa 250,000 a sakan daya. Ana iya saita ƙimar rx-μsecs-high mai zaman kanta daga rx-μsecs da tx-μsecs a cikin umarnin ethtool iri ɗaya, kuma ya kasance mai zaman kansa daga daidaitawar daidaitawar katsewa. Kayan aikin da ke ƙasa yana goyan bayan ƙima a cikin tazara na microsecond 2, don haka ƙimar maƙwabta na iya haifar da ƙimar katsewa iri ɗaya.

  • Don kashe daidaitawar katsewar daidaitawa: ethtool -C ethX adaptive-rx off adaptive-tx off
  • Don kunna daidaitawar katsewa na daidaitawa: ethtool -C ethX adaptive-rx akan adaptive-tx a kunne.

Kyakkyawan wuri don farawa don kunnawa gabaɗaya shine 84 μs, ko ~ 12000 katsewa/s. Idan ka ga rx_dropped counters suna gudana yayin zirga-zirga (ta amfani da ethtool -S ethX) to tabbas kuna da jinkirin CPU, rashin isassun buffers daga girman zoben adaftar (ethtool -G) don riƙe fakiti na 84 μs ko zuwa ƙarancin katsewa.

  • Don saita matsakaicin katsewa zuwa ƙayyadadden ƙimar katsewa na 84 μs tsakanin katsewa (12000 katsewa/s): ethtool -C ethX adaptive-rx off adaptive-tx off rx-usecs 84 tx-usecs 84 Ƙimar ta gaba da za a gwada, idan ba a ƙara girman ku akan amfani da CPU ba, shine 62 μs. Wannan yana amfani da ƙarin CPU, amma sabis yana buffers da sauri, kuma yana buƙatar ƴan sifofi (girman zobe, ethtool -G).
  • Don saita matsakaicin katsewa zuwa ƙayyadadden ƙimar katsewa na 62 usecs tsakanin katsewa (16000 katsewa/s). ethtool -C ethX adaptive-rx kashe adaptive-tx kashe rx-usecs 62 tx-usecs 62
    Idan rx_dropped ƙididdiga ta ƙaru yayin zirga-zirga (ta amfani da ethtool -S ethX), mai yiwuwa kuna da jinkirin CPU, rashin isassun abubuwan buffer daga girman zoben adaftan (ethtool -G), ko kuma ƙasa da ƙimar katsewa. Idan ba a fi ƙarfin ku akan amfani da CPU ba, zaku iya ƙara ƙimar katsewa ta rage ƙimar ITR. Wannan yana amfani da ƙarin CPU, amma sabis na buffer da sauri, kuma yana buƙatar ƙarancin siffantawa (girman zobe, ethtool -G).
    Idan CPU ɗinku yana kan 100%, to ba a ba da shawarar haɓaka ƙimar katsewa ba. A wasu yanayi kamar nauyin aikin da aka ɗaure na CPU, kuna iya ƙara ƙimar μs don kunna ƙarin lokacin CPU don wasu aikace-aikacen.
    Idan kuna buƙatar ƙarancin jinkirin aiki da/ko kuna da yawancin CPU don sadaukar da aikin hanyar sadarwa, zaku iya kashe matsakaicin katse gaba ɗaya, wanda ke ba da damar katse wuta da sauri.
  • Don kashe katse daidaitawar ethtool -C ethX adaptive-rx kashe adaptive-tx kashe rx-usecs 0 tx-usecs 0

NOTE
Lokacin da aka kashe gudu tare da katsewa, ƙimar katsewa akan kowane jerin gwano na iya zama babba. Yi la'akari da haɗa ma'aunin rx-usec-high don saita iyaka babba akan ƙimar katsewa. Umurnin da ke biyowa yana hana daidaitawar katsewar daidaitawa kuma yana ba da damar iyakar 5 micro seconds kafin ya nuna an kammala karɓa ko aikawa. Maimakon haifar da katsewa sama da 200,000 a cikin daƙiƙa guda, yana iyakance jimlar katsewa daƙiƙa guda zuwa 50,000 ta hanyar ma'aunin rx-usec-high. # ethtool -C ethX adaptive-rx off adaptive-tx off rx-usecs-high 20 rx-usecs 5 txusecs 5 Gwada daidaitawa da watsawa/karba/mafificin mahimmin lokaci coalescing mai girma (80/100/150/200) ko ƙasa (25/20/10) don zaɓar aikin aiki.
4.4 Girman zobe
Idan kuna ganin kirga rx_dropped a cikin ethtool -S ethX (rx_dropped, rx_dropped.nic), ko kuma ana zargin matsa lamba tare da layukan da yawa suna aiki, zaku iya gwada daidaita girman zobe daga ƙimar tsoho. Matsakaicin ƙimar shine 512, max shine 4096.

  • Don duba ƙimar halin yanzu: ethtool -g ethX
    Idan ana zargin rashin buffering yana haifar da faɗuwa a ƙimar katsewa na yanzu, zaku iya gwada matsakaicin farko, sannan mafi ƙaranci, sannan ku ci gaba a cikin binciken binary har sai kun ga kyakkyawan aiki.
    Idan ana zargin matsa lamba na cache (layi da yawa suna aiki) rage masu buffer daga tsoho na iya taimakawa Intel ® Data Direct I/O (Intel DDIO) yayi aiki tare da inganci. Intel yana ba da shawarar gwada 128 ko 256 a kowane jerin gwano, sanin cewa haɓaka ƙimar katsewa ta hanyar ethtool -C na iya zama dole don guje wa haɓakar rx_dropped.
  • Don saita girman zobe zuwa ƙayyadaddun ƙima: ethtool -G eth12 rx 256 tx 256

NOTE
Don gyara fakitin fakitin Rx da aka samo tare da ethtool -S ethX | grep digo, la'akari da ƙara girman zobe zuwa 4096. Gwaji don nemo mafi kyawun saiti don nauyin aiki amma kula da yawan amfani da ƙwaƙwalwar ajiya tare da ƙima mafi girma.
4.5 Gudanar da Yawo
Kula da kwararar Layer 2 na iya yin tasiri sosai akan aikin TCP kuma ana ba da shawarar a kashe shi don yawancin ayyukan aiki. Wata yuwuwar keɓanta shine fashe zirga-zirga inda fashewar ba ta daɗe ba.
An kashe sarrafa kwarara ta tsohuwa.

  • Don kunna sarrafa kwarara: ethtool -A ethX rx akan tx akan
  • Don musaki sarrafa kwarara: ethtool -A ethX rx kashe tx kashe

NOTE
Dole ne ku sami abokin haɗin haɗin gwiwa mai ikon sarrafa kwarara don samun nasarar kunna sarrafa kwarara.
4.6 Jumbo Frames
Lokacin da yanayin zirga-zirgar da ake tsammanin ya ƙunshi manyan tubalan bayanai da ake canjawa wuri, yana iya zama da amfani don kunna fasalin firam ɗin jumbo. Ana kunna goyan bayan Jumbo Frames ta hanyar canza Maximum Transmission Unit (MTU) zuwa ƙima mafi girma fiye da tsohuwar ƙimar 1500. Wannan yana bawa na'urar damar canja wurin bayanai a cikin manyan fakiti a cikin mahallin cibiyar sadarwa. Wannan saitin zai iya inganta kayan aiki da kuma rage amfani da CPU don manyan ayyukan I/O. Koyaya, yana iya yin tasiri ga ƙananan fakiti ko ayyukan aiki masu lanƙwasa.
NOTE
Firam ɗin Jumbo ko mafi girman saitin MTU dole ne a daidaita su yadda ya kamata a cikin mahallin cibiyar sadarwar ku.
Yi amfani da umarnin ifconfig don ƙara girman MTU. Domin misaliample, shigar da wadannan, inda shine lambar dubawa: ifconfig mtu 9000 up
A madadin, zaku iya amfani da umarnin ip kamar haka: ip link set mtu 9000 dev ip link kafa dev

Tuning Platform (I40e Ba Takamaiman ba)

5.1 Saitunan BIOS

  • Kunna Intel® VT-d don haɓaka aikin haɓakawa.
  • Hyper-threading (masu sarrafa ma'ana) na iya shafar aiki. Gwada shi a kunne ko a kashe don yawan aikin ku.
  • Intel® Turbo Boost yana ba da damar muryoyin CPU suyi aiki a mitoci mafi girma fiye da mitar tushe na CPU. Ƙaddamar da Intel® Turbo Boost zai iya inganta aiki don yawancin ayyuka masu yawa amma yana cin ƙarin iko don kiyaye kullun a mafi girma mita. Gwaji tare da kashe / kunna Turbo Boost don aikin ku.

NOTE
Ba a da garantin mitocin Turbo idan dandamali yana fuskantar babban amfani da CPU gabaɗaya. An rage mitocin turbo mafi girma yayin da yawan amfani da CPU ke ƙaruwa.
5.2 Gudanar da Wuta
Gudanar da wutar lantarki na iya yin tasiri ga aiki, musamman a cikin ƙananan ƙarancin aiki. Idan aikin yana da fifiko mafi girma fiye da rage yawan amfani da wutar lantarki, Intel yana ba da shawarar cewa kayi gwaji tare da iyakance tasirin sarrafa wutar lantarki. Akwai hanyoyi daban-daban don iyakance sarrafa wutar lantarki, ta hanyar kayan aikin tsarin aiki, saitunan BIOS, da sigogin taya na kwaya. Zaɓi hanya mafi kyau da matakin don dacewa da yanayin ku.
5.2.1 C-Sarkin Jiha
Ƙayyadade shigarwar C-jihar zuwa CO ko C1 yana inganta aiki kuma yana ƙara amfani da wutar lantarki.
Kashe CPU Package C6 shigarwar jihar na iya inganta aikin cibiyar sadarwa. Koyaya, wannan yana ƙara amfani da wutar lantarki.
Akwai zaɓuɓɓuka masu zuwa:

  • Mai ƙarfi sarrafa shigarwar jihar C:
    Bude
    /dev/cpu_dma_latency kuma rubuta iyakar latency da aka yarda dashi.

NOTE
Akwai ƙaramin shirin da ake kira cpudmalatency.c wanda za'a iya saukewa daga buɗaɗɗen tushen jama'a, harhadawa, kuma a gudanar da shi daga layin umarni don yin daidai wannan.
Mai zuwa example yana ba da damar μs biyar na lokacin tashi, don haka yana ba da damar shigar da C1: cpudmalatency 5 &

  • Iyakance iyakar C-state a cikin saitunan taya kernel:
    Don Intel CPUs: intel_idle.max_cstates=1
    Don CPUs marasa Intel: processor.max_cstates=1
  • Yi amfani da umarnin cpupower don dubawa da kashe jihar CPU C6: Duba: cpupower Monitor ko cpupower idle-info
    Kashe C6: cpupower mara amfani -d3 ko
    Kashe Jihohin C-Cpupower mara amfani -D0

Bayanan kula:

  1. Kashe jihohin C akan CPU idan uwar garken tana da Intel® 4th Gen Intel® Xeon® Scalable Processor(s). Lokacin da aka kunna ko kashe Hyper Threading, kashe jahohi marasa aiki (-D0) yana hana cibiyoyi shiga jihohi masu ƙarancin ƙarfi yayin lokutan aiki kuma yana rage jinkirin CPU don canzawa tsakanin jahohi marasa aiki da aiki.
  2. Gudanar da wutar lantarki na Intel® 4th Gen Intel® Xeon® Scalable Processor yana da tsauri sosai. Don kauce wa shigar da ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙananan ƙwayoyin da ake amfani da su don kiyaye su a farke na tsawon lokaci (ethtool -L). hade ). Hakanan, ɗaure katsewa zuwa takamaiman muryoyin ta amfani da saita irq affinity (mafi yawan lokuta tare da -x gida ko jerin abubuwan CPU), kuma tabbatar da cewa aikin yana gudana akan waɗancan nau'ikan nau'ikan guda ɗaya tare da saiti ko numactl. Wannan yana haɓaka aiki ta hanyar kiyaye muryoyin aiki da haɓaka sarrafa katsewa.

Kunna C6:
cpupower mara aiki -d3
Kunna Jihohin C:
cpupower rashin aiki -E

  • Wata hanya ita ce a yi amfani da kayan aikin da aka kunna (haɗe tare da yawancin rarrabawar Linux) don saita aikin profile. Wadannan profiles canza saitunan OS da yawa waɗanda zasu iya shafar aiki a cikin aikace-aikace da yawa. An gano cewa cibiyar sadarwa-throughput profile yana ba da haɓaka ga yawancin ayyukan aiki.
    Duba:
    tuned-adm aiki
    Saita:
    tuned-adm profile hanyar sadarwa ta hanyar sadarwa
    NOTE
    Sabis ɗin da aka kunna dole ne ya kasance yana gudana don umarni na sama. Don duba/sake kunnawa, kunna: systemctl yanayin da aka kunna systemctl sake kunnawa
    Hakanan zaka iya hana kowane shigarwar jihar C ta ƙara mai zuwa zuwa layin taya kernel:
    aiki = zabe
  • Iyakance jihar C ta tsarin saitin sarrafa wutar lantarki na BIOS, wanda zai iya samun ingantaccen aikinfile samuwa.
    Ana iya amfani da kayan aiki irin su turbostat ko x86_energy_perf_policy don dubawa ko saita saitunan sarrafa wutar lantarki.

5.2.2 PCIe Power Management
Gudanar da Wutar Lantarki na Jiha (ASPM) yana ba da damar ƙarancin wutar lantarki don hanyoyin haɗin PCIe lokacin da ba sa cikin amfani. Wannan na iya haifar da latency mai girma akan na'urorin cibiyar sadarwa na PCIe, don haka Intel ya ba da shawarar cewa ku kashe ASPM don ayyukan aiki masu latency. Kashe ASPM ta ƙara mai zuwa zuwa layin taya kernel: pcie_aspm=kashe
5.2.3 Matsakaicin Mitar CPU
Siffar mitar CPU (ko sikelin saurin CPU) dabara ce ta sarrafa wutar lantarki ta Linux wacce ake daidaita saurin agogon tsarin akan tashi don adana wuta da zafi. Kamar jihohin C, wannan na iya haifar da jinkirin da ba'a so akan haɗin yanar gizo.
Hakanan za'a iya amfani da kayan aikin cpupower don dubawa da gyara kuskuren aikin CPU da iyakoki:

  • Duba: cpupower Monitor ko
  • Saita CPUs zuwa yanayin aiki: aikin mitar cpupower-set-g

NOTE
Canje-canje zuwa iyakokin mitar CPU na iya yin tasiri ga yawancin ayyukan aiki kuma yana iya kashe wasu fasalulluka, kamar yanayin turbo na CPU.
Don musaki sikelin mitar CPU, kashe sabis na wutar lantarki ta hanyar umarni masu zuwa:
systemctl tasha cpupower.service
systemctl kashe cpupower.service
5.2.4 Ƙarin Jagorar Gudanar da Wuta
Ana ba da ƙarin cikakkun bayanai a cikin wannan babban matakinview na yawancin fasalulluka na sarrafa wutar lantarki a cikin na'urori masu sarrafawa na 3rd Generation Intel® Xeon® Scalable, da kuma jagora kan yadda za'a iya haɗa waɗannan fasalulluka a matakin dandamali: https://networkbuilders.intel.com/solutionslibrary/power-management-technologyoverview-technology-guide
5.3 Intel® Turbo Boost
Intel® Turbo Boost yana sa na'ura mai sarrafa sauri sauri lokacin da ake buƙata amma yana iya cinye ƙarin ƙarfi. Kashe Turbo Boost yana kiyaye na'ura mai sarrafawa a tsayin tsayin daka, yana ba ku daidaiton matakin aiki don takamaiman kayan aiki.
5.4 Firewalls
Firewalls na iya yin tasiri ga aiki, musamman aikin latency.
Kashe iptables/firewalld idan ba a buƙata ba.
5.5 Saitunan Aikace-aikace
Sau da yawa zaren guda ɗaya (wanda ya dace da layin hanyar sadarwa guda ɗaya) bai isa ya cimma matsakaicin bandwidth ba. Wasu gine-ginen dandali, kamar AMD, suna son sauke ƙarin fakitin Rx tare da zaren guda ɗaya idan aka kwatanta da dandamali tare da na'urori masu sarrafawa na Intel.
Yi la'akari da yin amfani da kayan aiki kamar saitin ɗawainiya ko numactl don haɗa aikace-aikace zuwa kumburin NUMA ko muryoyin CPU na gida zuwa na'urar cibiyar sadarwa. Don wasu nauyin aiki kamar ajiya I/O, matsar da aikace-aikacen zuwa kumburin da ba na cikin gida yana ba da fa'ida.
Gwaji tare da ƙara yawan zaren da aikace-aikacenku ke amfani da su idan zai yiwu.
5.6 Tsarin Kernel
Yawancin kernels na zamani na zamani an inganta su da kyau don aiki amma, dangane da yanayin amfani da ku, sabunta kwaya na iya samar da ingantaccen aiki. Zazzage tushen kuma yana ba ku damar kunna / kashe wasu fasalulluka kafin gina kernel.
5.7 Tsarin Aiki/Saitunan Kernel
Tuntuɓi jagororin daidaita tsarin aiki, kamar Jagoran Tunawa da Aiki na hanyar sadarwa na Red Hat Enterprise Linux, don ƙarin haske kan daidaita tsarin aiki gabaɗaya.
An jera wasu sigogi gama gari don kunnawa a cikin tebur mai zuwa. Lura cewa waɗannan wuraren farawa ne kawai aka ba da shawarar, kuma canza su daga abubuwan da ba a so ba na iya ƙara albarkatun da ake amfani da su akan tsarin. Kodayake haɓaka dabi'u na iya taimakawa wajen haɓaka aiki, ya zama dole don gwaji tare da ƙima daban-daban don sanin abin da ke aiki mafi kyau ga tsarin da aka ba da, aikin aiki da nau'in zirga-zirga.
Ana iya daidaita sigogin kernel ta amfani da sysctl mai amfani a cikin Linux kamar yadda aka nuna a ƙasa.
Zuwa view Tsofaffin dabi'u na rmem da wmem akan tsarin:
sysctl net.core.rmem_default
sysctl net.core.wmem_default
Saita ƙimar zuwa max (16 MB):
sysctl -w net.core.rmem_max=16777216
sysctl -w net.core.wmem_max=16777216
Girman buffer socket, wanda kuma aka sani da karɓar buffer (rmem) da watsa buffer (wmem), sigogi ne na tsarin da ke ƙayyadaddun adadin ƙwaƙwalwar ajiya da aka tanadar don zirga-zirgar hanyar sadarwa mai shigowa da mai fita.
Gudun sysctl ba tare da gardamar -w ba yana lissafin siga tare da saitin sa na yanzu.

Saitin Tari Bayani
net.core.rmem_default Tsohuwar Karɓa Girman Taga
net.core.wmem_default Tsohuwar Girman Taga mai watsawa
net.core.rmem_max Matsakaicin Girman Tagar Karɓa
net.core.wmem_max Matsakaicin Girman Taga mai watsawa
net.core.optmem_max Matsakaicin Matsalolin Ƙwaƙwalwar Ƙwaƙwalwa
net.core.netdev_max_backlog Bayanan fakitin da ba a sarrafa su ba kafin kernel ya fara raguwa
net.ipv4.tcp_rmem Mai ajiyar ƙwaƙwalwar ajiya don masu karanta TCP
net.ipv4.tcp_wmem Mai ajiyar ƙwaƙwalwar ajiya don TCP aika buffers

Kernel, tarin cibiyar sadarwa, mai sarrafa ƙwaƙwalwar ajiya, saurin CPU, da sigogin sarrafa wutar lantarki na iya yin babban tasiri akan aikin cibiyar sadarwa. Shawarar gama gari ita ce a yi amfani da shi zuwa ga kayan aikin hanyar sadarwafile ta amfani da umarnin da aka kunna. Wannan yana canza wasu saitunan OS don samar da fifiko ga aikace-aikacen sadarwar.
Duba:
tuned-adm aiki
Saita:
tuned-adm profile hanyar sadarwa ta hanyar sadarwa
5.8 Bayan Na'urar hanyar sadarwa
Wannan fasalin yana taimakawa haɓaka aikin cibiyar sadarwa ta hanyar sarrafa zirga-zirga masu shigowa yadda ya kamata, rage asarar fakiti, rage jinkiri, da haɓaka kayan aiki. Wannan yana haifar da mafi kyawun ƙwarewar mai amfani da amsawar tsarin sauri.
Ta hanyar tsoho, ana kunna shi a yawancin tsarin aiki na Linux. Don duba ƙimar tsoho:
sysctl net.core.netdev_max_backlog
Matsakaicin ƙimar netdev_max_backlog na iya bambanta dangane da abubuwa kamar sigar kernel, hardware, ƙwaƙwalwar ajiya, da nauyin aiki. A yawancin lokuta, ana ganin 8192 a matsayin darajar mai kyau. sysctl -w net.core.netdev_max_backlog=8192
5.9 Takamaiman Kanfigareshan da Tunatarwa
5.9.1 4th Generation Intel® Xeon® Scalable Processors
Gudanar da wutar lantarki na Intel® 4th Generation Intel® Xeon® Scalable processor yana da tsauri sosai idan aka kwatanta da na'urori masu sarrafawa na Intel® Xeon® Scalable na ƙarni na uku. Don kauce wa shiga cikin jahohi masu ƙarancin ƙarfi, gwada rage adadin muryoyin da ake amfani da su don kiyaye su a farke na tsawon lokaci.
Shawarar Saitunan Bios don Mafi Girman Aiki

  1. Ƙaddamar da zaren hyper-threading yana ba da damar / musaki (dangane da buƙatun aikin aiki da burin aiki) akan CPU.
  2. Saita tsarin profile zuwa Performance don matsakaicin aiki.
    NOTE
    Wannan yana haifar da yawan amfani da wutar lantarki
  3. Saita sarrafa wutar lantarki na CPU zuwa Mafi girman Aiki don ba da fifiko mafi girman aikin CPU akan ingancin wutar lantarki.
  4. Kunna Turbo Boost. Kashe Turbo Boost a cikin saitunan BIOS yawanci yana hana CPU haɓaka saurin agogo fiye da mitar sa.
  5. NOTE
    Kashe Turbo Boost na iya dacewa da wasu lokuta na amfani inda aka fifita daidaiton aiki, ingancin wutar lantarki, ko sarrafa zafi sama da mafi girman aiki.
  6. Kashe Siffar Tushen I/O Virtualization (SR-IOV), idan tsarin baya amfani da fasahohin haɓakawa.
  7. Kashe jihohin C don umurci CPU su ci gaba da aiki kuma su hana shiga cikin jahohi marasa aiki masu zurfi.
  8. Kashe C1E, don tabbatar da cewa CPU ya ci gaba da aiki kuma baya shigar da C1E mara amfani.
  9. Saita mitar uncore zuwa matsakaicin don ba da umarnin tsarin yin aiki a mafi girman mitar da ake samu.
  10. A kan dandamali na Dell, saita Teburin Bayanin APIC da yawa (MADT) ainihin kwaikwaya zuwa Linear (ko Round-Robin dangane da BIOS) don samar da taswirar bayyananniyar taswirar CPU.

Shawarwari na Tuna Matakan OS don Ingantattun Ayyuka

  1. Saita gwaman mitar mitar CPU zuwa aiki. saitin mitar cpupower -g aikin mita-info na cpupower
  2. Kashe Jihohin C. cpupower mara amfani -D0
  3. Saita ainihin Rx (rmem) da Tx (wmem) masu buffer zuwa iyakar ƙima. sysctl -w net.core.rmem_max=16777216 sysctl -w net.core.wmem_max=16777216
  4. Saita bayanan na'urar cibiyar sadarwa. sysctl -w net.core.netdev_max_backlog=8192
  5. Saita mai kunnawa profile (nauyin aiki ya dogara ga kayan aiki / latency).
    tuned-adm profile hanyar sadarwa ta hanyar sadarwa

Nasihar Tunawa Matakan Adafta don Ingantattun Ayyuka

  1. Iyakance adadin layukan da za a yi amfani da su don zirga-zirgar aikace-aikacen. Yi amfani da mafi ƙarancin adadin layukan da ake buƙata don kiyaye abubuwan haɗin CPU masu alaƙa suna aiki don hana su shiga cikin manyan jahohi marasa aiki (daidaita don aikin): ethtool -L hade 32
  2. Saita ƙimar matsakaicin katsewa. ethtool -C adaptive-rx kashe adaptive-tx kashe rx-usecs-high 50 rx-usecs 50 tx-usecs 50
    Gwada daidaita lokacin watsawa/ karɓa/mafi fifikon haɗakar lokaci mafi girma (80/100/150/200) ko ƙasa (25/20/10/5) don nemo madaidaicin ƙimar aikin.
  3. Saita girman zoben Rx/Tx. ethtool -G rx 4096 tx 4096
    NOTE
    Idan kun ga fakitin Rx ya sauke tare da ethtool -S| grep drop, gwada rage girman zobe zuwa <4096. Yi ƙoƙarin nemo mafi kyawun ƙimar aikin aiki inda ba a sauke fakiti ba.
  4. Saita IRQ Affinity. Yi amfani da cores na gida zuwa NIC, ko takamaiman taswirar taswira (inda # cores yayi daidai da adadin layukan da aka saita a shafi na 1 akan shafi na 26. systemctl stop irqbalance set_irq_affinity -X local OR saita_irq_affinity -X

5.9.2 AMD EPYC
Masu sarrafawa na AMD EPYC CPUs ne masu ƙarfi waɗanda aka yi don sabobin da cibiyoyin bayanai, waɗanda aka gina akan gine-ginen Zen na AMD. Saitunan da ke ƙasa sun fito ne daga jerin EPYC na ƙarni na 4 na AMD.
Nasihar Saitunan BIOS don Mafi Girman Aiki

  1. Kunna yanayin al'ada don ƙyale masu amfani su daidaita aikin CPU, amfani da wutar lantarki, da sauran saitunan. Wannan yana taimakawa wajen daidaita tsarin don mafi kyawun daidaito tsakanin aiki da ingantaccen makamashi.
  2. Ba da damar haɓaka aikin haɓakawa don ƙyale CPU ta ƙara saurinsa ta atomatik don ɗaukar ayyuka masu ƙarfi, haɓaka aikin gabaɗaya.
  3. Kashe iko na duniya na C-state, don hana CPU shiga cikin zurfin ikon ceton jihohin da aka sani da jihohin C, wanda zai iya kula da amsa.
    NOTE
    Kashe jihohin C na iya haifar da ƙarin amfani da wutar lantarki da haɓaka yanayin zafi. Saka idanu biyu don nauyin aiki.
  4. Kunna/ kashe Multithreading na lokaci ɗaya (SMT) akan CPU, dangane da buƙatun aikin aiki da burin aiki. SMT yayi daidai da Hyper Threading akan Intel CPUs.
    NOTE
    Don ingantaccen aiki, koma zuwa Tuning i40e Direba Saituna a shafi na 13 da Platform Tuning (i40e Non-Specific) a shafi na 19 don ingantaccen OS da daidaita matakin adafta.

Adafta Bonding

Haɗin Linux wani abu ne mai ƙarfi wanda zai iya inganta aikin cibiyar sadarwa, sakewa, da haƙurin kuskure a cikin mahallin uwar garken. Koyaya, yana da mahimmanci a lura cewa yana buƙatar kayan aikin cibiyar sadarwa masu dacewa da ingantaccen tsari akan uwar garken da maɓalli don aiki yadda yakamata.
Direban haɗin kai a cikin Linux yana ba ku damar haɗa mahaɗar hanyoyin sadarwa na zahiri da yawa zuwa haɗin haɗin gwiwa. Wannan haɗin haɗin yanar gizon yana bayyana azaman hanyar sadarwa ta hanyar sadarwa guda ɗaya zuwa tsarin aiki da aikace-aikace.
NOTE
Haɗin yana da ma'ana mai ma'ana, don haka ba zai yiwu a saita alaƙar CPU kai tsaye akan haɗin haɗin gwiwa ba (misaliample, bond0). Wato, ba ta da iko kai tsaye kan sarrafa katse ko alaƙar CPU. Dole ne a saita alaƙar CPU don madaidaitan mu'amala waɗanda ke ɓangaren haɗin gwiwa.
Bonding yana ba da nau'ikan ayyuka da yawa, kowanne yana da halayensa.

Yanayin  Nau'in
0 Zagaye Robin
1 Ajiyayyen Aiki
2 XOR
3 Watsa shirye-shirye
4 LACP
5 Watsa Ma'aunin Load
6 Ma'aunin Load Mai Adaɗi

Akwai hanyoyi daban-daban don ƙirƙirar haɗin gwiwa a cikin Linux. Ɗaya daga cikin hanyoyin gama gari ita ce ta amfani da saitin hanyar sadarwa files (kamar example, /etc/network/ musaya ko /etc/sysconfig/network-scripts/ifcfg-bondX).
Kanfigareshan Amfani da Kanfigareshan hanyar sadarwa Files
Matakan da ke biyowa suna haifar da haɗin kai ta hanyar saitin hanyar sadarwa files.

  1. Zaɓi tashoshin NIC biyu ko fiye don haɗin gwiwa (misaliample, ethX da ethY)
  2. Bude Kanfigareshan NIC Files karkashin /etc/sysconfig/network-scripts/ don Interface NIC da ake buƙata (misaliample, vi ifcfg-ethX da vi ifcfg-ethY) kuma saka rubutu mai zuwa:
    MASTER=bondN [Lura: N shine lamba don ambaton lambar haɗin.] BAYI=ee.
  3. Ƙirƙiri rubutun haɗin gwiwa file ta amfani da vi /etc/sysconfig/networkscripts/ifcfg-bondN kuma shigar da rubutu mai zuwa:
    Na'ura = bondN [Lura: N shine lamba don ambaton lambar haɗin] ONBOOT = eh USERCTL = babu BOOTPROTO =dhcp (ko) babu
    IPADDR=200.20.2.4 [ana bukatar idan BOOTPROTO=babu] NETMASK=255.255.255.0 BOOTPROTO=babu] BONDING_OPTS=”yanayin=200.20.2.0 miimon=200.20.2.255″
    NOTE
    Yanayin zai iya zama kowace lamba daga 0 zuwa 6 bisa ga abin da ake bukata.
  4. Sake kunna sabis na cibiyar sadarwa ta amfani da cibiyar sadarwar sabis zata sake farawa ko systemctl zata sake farawa NetworkManager.service

Magance Matsalar Aiki

7.1 Amfani da CPU
Bincika amfani da CPU a kowane tushe yayin da aikin ke gudana.
Yi la'akari da cewa yin amfani da kowane tushe ya fi dacewa da aiki fiye da amfani da CPU gabaɗaya tunda yana ba da ra'ayin amfani da CPU kowane layin hanyar sadarwa. Idan kuna da ƴan zaren da ke tafiyar da zirga-zirgar hanyar sadarwa, to za ku iya samun ƴan muryoyin da ake amfani da su. Koyaya, idan waɗannan muryoyin suna cikin 100%, to ana iya iyakance kayan aikin hanyar sadarwar ku ta amfani da CPU kuma lokaci yayi da za a yi masu zuwa:

  1. Sake daidaita girman zobe na IRQ kamar yadda aka yi dalla-dalla a cikin Tsayar da Tsagewar.
  2. Ƙara adadin zaren aikace-aikacen don yada nauyin CPU akan ƙarin muryoyin. Idan duk muryoyin suna gudana a 100% to aikace-aikacen ku na iya zama daure CPU maimakon hanyar sadarwa.

Kayan aikin da aka saba samu:

  • saman
    - Latsa 1 don faɗaɗa jerin CPUs kuma bincika waɗanda ake amfani da su.
    - Lura da matakin amfani.
    - Yi la'akari da waɗanne matakai aka jera a matsayin mafi yawan aiki (saman lissafin).
  • mpstat
    Mai zuwa exampAn gwada layin umarni akan Red Hat Enterprise Linux 7.x.
    Yana nuna amfani da CPU akan kowane tushe (ta hanyar nemo jimlar kashi mara amfani da raguwa daga 100) kuma yana nuna ƙimar sama da 80% a ja. mpstat -P DUK 1 1 | grep -v Matsakaici | wutsiya -n +5 | kai -n -1 | awk '{buga (100-$13)}' | egrep -color=ko da yaushe '[^\.][8-9][0-9][\.]?.*|^[8-9][0-9][\.]?.*| 100|' | shafi
  • perf top Nemo inda ake kashe keken keke.

7.2 i40e Counters
Direban i40e yana ba da jerin jerin ƙididdiga masu tsayi don gyara kurakurai da saka idanu ta hanyar ethtool -S ethX umurnin. Zai iya zama taimako don kallon fitarwa yayin da nauyin aiki ke gudana da/ko kwatanta ƙimar ƙima kafin da bayan aikin aiki.

  • Don samun cikakken juji na i40e counters: ethtool -S ethX
  • Don kallon ƙididdiga marasa sifili kawai: watch -d (ethtool -S ethX) | egrep -v :\ 0 | shafi
    Wasu abubuwan da ake nema:
  • rx_dropped yana nufin CPU baya yin hidimar buffers cikin sauri.
  • port.rx_dropped yana nufin wani abu baya da sauri a cikin ramin / ƙwaƙwalwar / tsarin.

7.3 Cibiyar sadarwa Counters
Bincika netstat-s kafin/bayan aikin aikin.
Netstat yana tattara bayanan cibiyar sadarwa daga duk na'urorin sadarwar da ke cikin tsarin. Don haka, ƙila za a iya yin tasiri ga sakamako daga cibiyoyin sadarwa ban da cibiyar sadarwar da ake gwadawa. Fitowa daga netstat -s na iya zama kyakkyawan nuni na al'amuran aiki a cikin tsarin aiki na Linux ko kwaya. Tuntuɓi jagororin daidaita tsarin aiki, kamar Jagoran Tunawa da Aiki na hanyar sadarwa na Red Hat Enterprise Linux, don ƙarin haske kan daidaita tsarin aiki gabaɗaya.
7.4 Lissafin tsarin
Bincika rajistan ayyukan don kurakurai da faɗakarwa (/var/log/messages, dmesg).
7.5 Intel svr-info Tool
Intel yana ba da kayan aikin svr-info (duba https://github.com/intel/svr-info) don Linux wanda ke ɗaukar bayanan hardware da software masu dacewa daga sabar. fitowar svr-info na iya zama da taimako sosai don gano ƙullawar tsarin ko saituna/ tuning waɗanda ba a inganta su don nauyin aiki ba. Lokacin buɗe shari'ar goyan baya tare da Intel don abubuwan da suka shafi aikin Ethernet, tabbas kun haɗa da fitowar bayanan svr (rubutu file) ga kowane uwar garken Linux a cikin tsarin gwaji.

  1. Zazzage kuma shigar da bayanan svr-info:
    wget -qO- https://github.com/intel/svr-info/releases/latest/download/svrinfo.tgz| tar xvz cd svr-info
    ./svr-info
    > hostname.txt
  2. Tattara abubuwan da aka fitar:
    ./svr-info > hostname.txt
  3. Haɗa rubutu ɗaya (.txt) file ga kowane uwar garken zuwa yanayin tallafin Intel don bincike.

Shawarwari don Yanayin Ayyuka na gama-gari

8.1 IP Gabatarwa

  • Sabunta kwaya.
    Wasu kwayayen in-distro na baya-bayan nan sun ɓata aikin tuƙi saboda sauye-sauyen kernel a lambar tuƙi wanda ya fara tare da cire cache ɗin tuƙi saboda tsaro. Kwanan baya-bayan kernels yakamata su sami faci waɗanda zasu rage tasirin waɗannan canje-canje kuma suna iya samar da ingantaccen aiki.
  • Kashe hyper-threading (maganin ma'ana).
  • Shirya sigogin taya na kwaya.
    - Kashe iommu (intel_iommu = kashe ko iommu = kashe) daga layin taya kernel sai dai idan an buƙata don haɓakawa.
    - Kashe sarrafa wutar lantarki: processor.max_cstates=1 aiki = zabe pcie_aspm = kashe
  • Ƙayyade adadin jerin layi don zama daidai da adadin muryoyi akan soket na gida (12 a cikin wannan tsohonample). ethtool -L ethX hade 12
  • Pin yana katsewa zuwa soket na gida kawai. set_irq_affinity -X gida ethX KO saita_irq_affinity -X gida ethX
    NOTE
    Za a iya amfani da -X ko -x dangane da nauyin aiki.
  • Canza girman zoben Tx da Rx kamar yadda ake buƙata. Ƙimar da ta fi girma tana ɗaukar ƙarin albarkatu amma tana iya samar da ƙimar isarwa mafi kyau. ethtool -G ethX rx 4096 tx 4096
  • Kashe GRO lokacin da ake tuƙi.
    Saboda sanannen batun kernel, GRO dole ne a kashe lokacin da ake turawa. ethtool -K ethX ya ƙare inda ethX shine keɓancewar Ethernet da za a gyara.
  • Kashe daidaitawar katsewa na daidaitawa kuma saita ƙima mai tsayi. ethtool -C ethX adaptive-rx kashe adaptive-tx kashe ethtool -C ethX rx-usecs 64 tx-usecs 64

NOTE
Dangane da nau'in sarrafawa da nauyin aiki, ana iya daidaita sigogin haɗakarwa na RX da TX don ingantaccen aiki (ko ƙasan asarar firam).

  • Kashe Tacewar zaɓi. sudo systemctl kashe firewalld sudo systemctl dakatar da firewalld
  • Kunna tura IP. sysctl -w net.ipv4.ip_forward=1
  • Tsara matsakaicin ƙima don karɓa kuma aika masu girman buffer soket. sysctl -w net.core.rmem_max=16777216 sysctl -w net.core.wmem_max=16777216

NOTE
Dangane da nauyin aiki ko buƙatu, ana iya canza waɗannan ƙimar daga tsoho.
8.2 Rashin Lantarki

  • Kashe hyper-threading (maganin ma'ana) KASHE.
  • Tabbatar cewa na'urar sadarwar tana gida zuwa numa core 0.
  • Sanya ma'auni zuwa ainihin 0 ta amfani da taskset-c 0.
  • Kashe irqbalance ta amfani da systemctl tasha irqbalance ko systemctl musaki irqbalance
  • Gudanar da rubutun alaƙa don yadawa a kan muryoyin. Gwada ko dai na gida ko duka.
  • Kashe daidaitawar katsewa. ethtool -C ethX rx-usecs 0 tx-usecs 0 adaptive-rx off adaptive-tx off rxusecs- high 0
  • Iyakance adadin layuka don zama daidai da adadin muryoyi akan soket na gida (32 a cikin wannan tsohonample). ethtool -L ethX hade 32
  • Pin yana katsewa zuwa soket na gida kawai (rubutun kunshe da tushen direban i40e). set_irq_affinity -X gida ethX
  • Yi amfani da kafaffen ma'auni kamar netperf -t TCP_RR, netperf -t UDP_RR, ko NetPipe. netperf -t TCP_RR ko netperf -t UDP_RR
  • Sanya ma'auni zuwa cibiya guda ɗaya a cikin kumburin NUMA na gida. aiki -c

Intel ® Ethernet 700 Series
Jagorar Tunawa Ayyukan Linux
Disamba 2024
Doc. No.: 334019, Rev.: 1.2

Takardu / Albarkatu

Intel Ethernet 700 Series Linux Performance Tuning [pdf] Jagorar mai amfani
334019, Ethernet 700 Series Linux Performance Tuning, Ethernet 700 Series, Linux Performance Tuning, Performance Tuning, Tuning

Magana

Bar sharhi

Ba za a buga adireshin imel ɗin ku ba. Ana yiwa filayen da ake buƙata alama *