Sanibonani nonke. Siyaqhubeka nokukwethula kusistimu yokugcina idatha ye-Aerodisk VOSTOK, esekelwe ku-Russian Elbrus 8C processor.
Kulesi sihloko thina (njengoba kuthenjisiwe) sizohlaziya ngokuningiliziwe esinye sezihloko ethandwa kakhulu futhi ezithakazelisayo ezihlobene Elbrus, okungukuthi ukukhiqiza. Kukhona ukuqagela okuningi ngokusebenza kwe-Elbrus, kanye ne-polar ngokuphelele. Ama-Pessimists athi ukukhiqiza kuka-Elbrus manje "akunalutho", futhi kuzothatha amashumi eminyaka ukubamba abakhiqizi "abaphezulu" (okungukuthi, eqinisweni lamanje, neze). Ngakolunye uhlangothi, abanethemba lokuthi i-Elbrus 8C isivele ibonisa imiphumela emihle, futhi eminyakeni embalwa ezayo, ngokukhululwa kwezinguqulo ezintsha zamaphrosesa (Elbrus 16C no-32C), sizokwazi "ukubamba futhi sidlule" abakhiqizi beprosesa abahamba phambili emhlabeni.
Thina kwa-Aerodisk singabantu abasebenzayo, ngakho-ke sithathe umzila olula noqondakala kakhulu (kithi): ukuhlola, ukurekhoda imiphumela bese sifinyelela iziphetho. Ngenxa yalokho, senze izivivinyo eziningi kakhulu futhi sathola izici zokusebenza eziningi ze-Elbrus 8C e2k architecture (kuhlanganise nezinhle) futhi, vele, saqhathanisa lokhu nezinhlelo zokulondoloza ezifanayo kuma-Intel Xeon amd64 architecture processors.
Kodwa-ke, sizokhuluma kabanzi mayelana nokuhlolwa, imiphumela kanye nokuthuthukiswa kwesikhathi esizayo kwezinhlelo zokugcina ku-Elbrus ku-webinar yethu elandelayo ethi "OkoloIT" ngo-Okthoba 15.10.2020, 15 ngo-00:XNUMX. Ungabhalisa usebenzisa isixhumanisi esingezansi.
→
Isitendi sokuhlola
Senze izitendi ezimbili. Zombili lezi zimo zineseva esebenzisa i-Linux, exhunywe ngokushintsha kwe-16G FC kwizilawuli ezimbili zokugcina, lapho kufakwa khona amadiski ayi-12 e-SAS SSD 960 GB (11,5 TB "yomthamo ongahluziwe" noma i-5,7 TB yomthamo "osebenzisekayo", uma sisebenzisa i-RAID. -10).
Ngokohlelo isitendi sibukeka kanje.
Isitendi esingu-1 e2k (Elbrus)
Ukucushwa kwe-hardware kungokulandelayo:
- Iseva ye-Linux (2xIntel Xeon E5-2603 v4 (6 cores, 1,70Ghz), 64 GB DDR4, 2xFC adaptha 16G 2 port) - 1 pc.
- Shintsha i-FC 16 G - 2 ama-pcs.
- Isitoreji uhlelo Aerodisk Vostok 2-E12 (2xElbrus 8C (8 cores, 1,20Ghz), 32 GB DDR3, 2xFE FC-adaptha 16G 2 port, 12xSAS SSD 960 GB) - 1 pc.
Isitendi esingu-2 amd64 (Intel)
Ukuze uqhathanise nokucushwa okufanayo ku-e2k, sisebenzise ukucushwa kwesitoreji okufanayo ngephrosesa efana nezici ku-amd64:
- Iseva ye-Linux (2xIntel Xeon E5-2603 v4 (6 cores, 1,70Ghz), 64 GB DDR4, 2xFC adaptha 16G 2 port) - 1 pc.
- Shintsha i-FC 16 G - 2 ama-pcs.
- Isitoreji uhlelo Aerodisk Engine N2 (2xIntel Xeon E5-2603 v4 (6 cores, 1,70Ghz), 32 GB DDR4, 2xFE FC-adaptha 16G 2 port, 12xSAS SSD 960 GB) - 1 pc.
Inothi elibalulekile: Amaphrosesa we-Elbrus 8C asetshenziswa ekuhlolweni asekela i-DDR3 RAM kuphela, lokhu "kubi, kodwa hhayi isikhathi eside." I-Elbrus 8SV (asinayo esitokweni okwamanje, kodwa sizoba nayo maduze) isekela i-DDR4.
Indlela yokuhlola
Ukuze sikhiqize umthwalo, sisebenzise uhlelo oludumile noluhlolwe isikhathi lwe-Flexible IO (FIO).
Zombili izinhlelo zokugcina zilungiswa ngokuvumelana nezincomo zethu zokucushwa, ngokusekelwe ezidingweni zokusebenza okuphezulu ekufinyeleleni kwebhulokhi, ngakho sisebenzisa amachibi ediski e-DDP (Dynamic Disk Pool). Ukuze singahlanekezeli imiphumela yokuhlolwa, sikhubaza ukuminyanisa, ukuphindaphinda kanye nenqolobane ye-RAM kuwo womabili amasistimu okugcina.
Ama-D-LUN angu-8 adalwe ku-RAID-10, 500 GB ngayinye, enenani eliphelele elisebenzisekayo le-4 TB (okungukuthi, cishe u-70% wamandla asebenzisekayo angenzeka walokhu kucushwa).
Izimo eziyisisekelo nezidumile zokusebenzisa amasistimu okugcina zizokwenziwa, ikakhulukazi:
izivivinyo ezimbili zokuqala zilingisa ukusebenza kwe-DBMS yokuthengiselana. Kuleli qembu lokuhlola sinentshisekelo ku-IOPS nokubambezeleka.
1) Ukufunda ngokungahleliwe emabhuloki amancane 4k
a. Usayizi wokuvimba = 4k
b. Funda/Bhala = 100%/0%
c. Inani lemisebenzi = 8
d. Ukujula komugqa = 32
e. Layisha uhlamvu = Okungahleliwe Okugcwele
2) Ukurekhoda okungahleliwe kumabhulokhi amancane 4k
a. Usayizi wokuvimba = 4k
b. Funda/Bhala = 0%/100%
c. Inani lemisebenzi = 8
d. Ukujula komugqa = 32
e. Layisha uhlamvu = Okungahleliwe Okugcwele
ukuhlolwa okubili kwesibili kulingisa ukusebenza kwengxenye yokuhlaziya ye-DBMS. Kuleli qembu lokuhlola sinentshisekelo ku-IOPS nokubambezeleka.
3) Ukufunda ngokulandelana emabhuloki amancane 4k
a. Usayizi wokuvimba = 4k
b. Funda/Bhala = 100%/0%
c. Inani lemisebenzi = 8
d. Ukujula komugqa = 32
e. Layisha uhlamvu = Okulandelanayo
4) Ukurekhoda okulandelanayo kumabhulokhi amancane 4k
a. Usayizi wokuvimba = 4k
b. Funda/Bhala = 0%/100%
c. Inani lemisebenzi = 8
d. Ukujula komugqa = 32
e. Layisha uhlamvu = Okulandelanayo
Iqembu lesithathu lokuhlola lilingisa umsebenzi wokufunda ukusakaza (isibonelo: ukusakazwa kwe-inthanethi, ukubuyisela izipele) nokurekhoda kokusakaza (isibonelo: ukubhekwa kwevidiyo, ukurekhoda izipele). Kuleli qembu lokuhlola, asisenayo intshisekelo ku-IOPS, kodwa ku-MB/s kanye nokubambezeleka.
5) Ukufunda ngokulandelana emabhuloki amakhulu angu-128k
a. Usayizi wokuvimba = 128k
b. Funda/Bhala = 0%/100%
c. Inani lemisebenzi = 8
d. Ukujula komugqa = 32
e. Layisha uhlamvu = Okulandelanayo
6) Ukurekhoda okulandelanayo kumabhulokhi amakhulu we-128k
a. Usayizi wokuvimba = 128k
b. Funda/Bhala = 0%/100%
c. Inani lemisebenzi = 8
d. Ukujula komugqa = 32
e. Layisha uhlamvu = Okulandelanayo
Ukuhlolwa ngakunye kuzohlala ihora elilodwa, kungafaki isikhathi sokufudumala samalungu afanayo amaminithi angu-7.
Imiphumela yokuhlola
Imiphumela yokuhlolwa ifingqwe ngamathebula amabili.
I-Elbrus 8S (SHD Aerodisk Vostok 2-E12)
I-Intel Xeon E5-2603 v4 (Isistimu yokugcina i-Aerodisk Engine N2)
Imiphumela yabonakala ijabulisa kakhulu. Kuzo zombili izimo, siwasebenzise kahle amandla okucubungula esistimu yesitoreji (ukusetshenziswa okungu-70-90%), futhi kulesi simo, ubuhle nobubi bawo womabili amaphrosesa kubonakala ngokucacile.
Kuwo womabili amathebula, ukuhlola lapho abacubungulayo “bezizwa bezethemba” futhi babonise imiphumela emihle kugqanyiswa ngokuluhlaza, kuyilapho izimo abacubunguli “abangazithandi” zigqanyiswa ngowolintshi.
Uma sikhuluma ngomthwalo ongahleliwe kumabhulokhi amancane, khona-ke:
- ngokombono wokufunda okungahleliwe, i-Intel ngokuqinisekile ingaphambi kuka-Elbrus, umehluko izikhathi ezingu-2;
- ngokombono wokuqopha okungahleliwe nakanjani kuwumdwebo, womabili amaphrosesa abonise imiphumela ecishe ilingane nehloniphekile.
Kumthwalo olandelanayo kumabhulokhi amancane isithombe sihlukile:
- kokubili lapho ufunda futhi ubhala, i-Intel ibaluleke kakhulu (izikhathi ezi-2) ngaphambi kuka-Elbrus. Ngesikhathi esifanayo, uma i-Elbrus inenkomba ye-IOPS ephansi kune-Intel, kodwa ibukeka ihloniphekile (izinkulungwane ezingama-200-300), khona-ke kunenkinga esobala yokubambezeleka (iphindwe kathathu kune-Intel). Isiphetho, inguqulo yamanje ye-Elbrus 8C "ayithandi" imithwalo elandelanayo kumabhulokhi amancane. Kuyacaca ukuthi kunomsebenzi okufanele wenziwe.
Kodwa ngomthwalo olandelanayo onamabhulokhi amakhulu, isithombe sihluke ngokuphelele:
- womabili amaphrosesa abonise cishe imiphumela elinganayo ku-MB/s, kodwa kukhona eyodwa KODWA.... Ukusebenza kwe-latency kuka-Elbrus kungcono izikhathi ezingu-10 (ishumi, Karl!!!) (okungukuthi ngaphansi) kunaleyo yephrosesa efanayo evela ku-Intel (0,4/0,5 ms iqhathaniswa no-5,1/6,5 ms) . Ekuqaleni besicabanga ukuthi yiphutha, ngakho-ke saphinda sahlola imiphumela, sahlola kabusha, kodwa esibuye sahlola sabonisa isithombe esifanayo. Lokhu kuyinzuzo engathi sína ye-Elbrus (kanye nesakhiwo se-e2k ngokujwayelekile) ngaphezu kwe-Intel (futhi, ngokufanelekile, nesakhiwo se-amd64). Asethembe ukuthi le mpumelelo izothuthukiswa.
Kukhona esinye isici esithakazelisayo sika-Elbrus, umfundi olalelayo angakwazi ukusinaka ngokubuka itafula. Uma ubheka umehluko phakathi kokusebenza kokufunda nokubhala kwe-Intel, kuzo zonke izivivinyo, ukufunda kungaphambi kokubhala ngokwesilinganiso cishe ngo-50%+. Lena inkambiso wonke umuntu (kuhlanganise nathi) ayijwayele. Uma ubheka i-Elbrus, izinkomba zokubhala ziseduze kakhulu nezinkomba zokufunda, ukufunda ngaphambi kokubhala, njengomthetho, ngo-10 - 30%, kungasekho.
Kusho ukuthini lokhu? Iqiniso lokuthi u-Elbrus "uthanda ngempela" ukubhala, futhi lokhu, futhi, kusikisela ukuthi le processor izoba usizo kakhulu emisebenzini lapho ukubhala kunqoba ngokucacile phezu kokufunda (ngubani owathi umthetho ka-Yarovaya?), Okubuye kube inzuzo engangabazeki ye-e2k architecture, futhi le nzuzo idinga ukuthuthukiswa.
Iziphetho kanye nekusasa eliseduze
Ukuhlolwa okuqhathanisayo kwe-Elbrus kanye ne-Intel ephakathi kwebanga eliphakathi kwemisebenzi yokugcina idatha kubonise cishe imiphumela elinganayo nefaneleka ngokulinganayo, kuyilapho iphrosesa ngayinye ibonise izici zayo ezithakazelisayo.
U-Intel waphumelela kakhulu u-Elbrus ekufundeni okungahleliwe kumabhulokhi amancane, kanye nasekufundeni okulandelanayo nokubhala ngamabhulokhi amancane.
Lapho ubhala ngokungahleliwe kumabhulokhi amancane, womabili amaphrosesa abonisa imiphumela elinganayo.
Ngokuya nge-latency, i-Elbrus ibukeka ingcono kakhulu kune-Intel ekulayisheni kokusakaza, i.e. ekufundeni nasekubhaleni ngokulandelana kwazo emabhuloki amakhulu.
Ngaphezu kwalokho, u-Elbrus, ngokungafani ne-Intel, ubhekana kahle ngokulinganayo kokubili ukufunda nokubhala imithwalo, kuyilapho nge-Intel, ukufunda kuhlale kungcono kakhulu kunokubhala.
Ngokusekelwe emiphumeleni etholiwe, singafinyelela isiphetho mayelana nokusetshenziswa kwezinhlelo zokugcina idatha ze-Aerodisk Vostok kuphrosesa ye-Elbrus 8C kule misebenzi elandelayo:
- amasistimu wolwazi anemisebenzi eminingi yokubhala;
- ukufinyelela ifayela;
- ukusakazwa kwe-inthanethi;
- I-CCTV;
- ikhophi yasenqolobaneni;
- okuqukethwe kwemidiya.
Ithimba le-MCST lisenokuthile okumele lisebenze, kodwa umphumela womsebenzi wabo usuvele ubonakala, okuyinto, ngokuqinisekile, ayikwazi kodwa ukuthokoza.
Lezi zivivinyo zenziwa ku-Linux kernel ye-e2k version 4.19; okwamanje ezivivinyweni ze-beta (ku-MCST, e-Basalt SPO, futhi lapha e-Aerodisk) kukhona i-Linux kernel 5.4-e2k, lapho, phakathi kwezinye izinto, iklanywe kabusha ishejuli kanye nokulungiselelwa okuningi kwamadrayivu e-solid state anesivinini esikhulu. Futhi, ikakhulukazi izikhwebu zegatsha le-5.x.x, i-MCST JSC ikhipha i-LCC compiler entsha, inguqulo 1.25. Ngokwemiphumela yokuqala, kuphrosesa efanayo ye-Elbrus 8C, i-kernel entsha ehlanganiswe ne-compiler entsha, imvelo ye-kernel, izinsiza zesistimu nemitapo yolwazi futhi, empeleni, isofthiwe ye-Aerodisk VOSTOK izovumela ukwanda okubaluleke nakakhulu ekusebenzeni. Futhi lokhu ngaphandle kokufaka imishini - kuphrosesa efanayo kanye namafrikhwensi afanayo.
Silindele ukukhishwa kwenguqulo ye-Aerodisk VOSTOK esekelwe ku-kernel 5.4 ngasekupheleni konyaka, futhi ngokushesha nje lapho umsebenzi wenguqulo entsha usuqediwe, sizobuyekeza imiphumela yokuhlolwa futhi siyishicilele lapha.
Uma manje sibuyela ekuqaleni kwesihloko bese siphendula umbuzo, ngubani oqinisile: abantu abangenathemba abathi u-Elbrus "ayilutho" futhi akasoze athola abakhiqizi abahamba phambili beprosesa, noma abanethemba lokuhle abathi "sebevele babanjwe. futhi uzodlula maduze"? Uma singaqhubeki ekucabangeni nasekubandlululeni ngokwenkolo, kodwa kusukela ekuhlolweni kwangempela, khona-ke abanethemba lokuhle baqinisile impela.
I-Elbrus isivele ikhombisa imiphumela emihle uma iqhathaniswa namaphrosesa we-amd64 asezingeni eliphakathi. I-Elbrus 8-ke, vele, ikude namamodeli aphezulu kakhulu we-server processors avela ku-Intel noma i-AMD, kodwa ibingahlosiwe lapho; ama-processor 16C kanye ne-32C azokhululwa ngale njongo. Bese sizokhuluma.
Siyaqonda ukuthi ngemva kwalesi sihloko kuzoba nemibuzo eyengeziwe mayelana ne-Elbrus, ngakho-ke sinqume ukuhlela enye i-webinar ye-inthanethi ethi "OkoloIT" ukuze siphendule le mibuzo bukhoma.
Kulokhu isivakashi sethu sizoba yiPhini Lomqondisi-Jikelele wenkampani ye-MCST, u-Konstantin Trushkin. Ungabhalisela i-webinar usebenzisa isixhumanisi esingezansi.
→
Siyabonga nonke, njengenjwayelo, sibheke ngabomvu ukugxeka okwakhayo nemibuzo ethokozisayo.
Source: www.habr.com