{"id":95783,"date":"2020-10-03T13:42:02","date_gmt":"2020-10-03T11:42:02","guid":{"rendered":"https:\/\/prohoster.info\/blog\/administrirovanie\/nastrojka-yadra-linux-dlya-glusterfs"},"modified":"2020-10-03T13:42:02","modified_gmt":"2020-10-03T11:42:02","slug":"nastrojka-yadra-linux-dlya-glusterfs","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/administrirovanie\/nastrojka-yadra-linux-dlya-glusterfs","title":{"rendered":"Konfigurimi i b\u00ebrtham\u00ebs Linux p\u00ebr GlusterFS","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

P\u00ebrkthimi i artikullit \u00ebsht\u00eb p\u00ebrgatitur n\u00eb prag t\u00eb fillimit t\u00eb kursit \u00abAdministrator Linux. Professional\u00bb<\/a><\/noindex>.<\/i><\/b><\/p>\n

\"Konfigurimi<\/p>\n

Her\u00eb pas here, shfaqen pyetje rreth rekomandimeve t\u00eb Gluster p\u00ebr konfigurimin e b\u00ebrtham\u00ebs dhe n\u00ebse ka ndonj\u00eb nevoj\u00eb p\u00ebr to.<\/p>\n

Nj\u00eb nevoj\u00eb e till\u00eb ndodhi rrall\u00eb. N\u00eb shumic\u00ebn e ngarkesave, b\u00ebrthama funksionon shum\u00eb mir\u00eb. Megjithat\u00eb, ka dhe nj\u00eb an\u00eb tjet\u00ebr. Historikisht, b\u00ebrthama Linux \u00ebsht\u00eb e gatshme t\u00eb konsumoj\u00eb shum\u00eb memorie, n\u00ebse i jepet mund\u00ebsia, p\u00ebrfshir\u00eb p\u00ebr memorizimin si m\u00ebnyra kryesore p\u00ebr t\u00eb p\u00ebrmir\u00ebsuar performanc\u00ebn.<\/a><\/noindex><\/p>\n

N\u00eb shumic\u00ebn e rasteve, kjo funksionon shk\u00eblqyesh\u00ebm, por n\u00eb ngarkesa t\u00eb m\u00ebdha mund t\u00eb \u00e7oj\u00eb n\u00eb probleme. <\/p>\n

Ne kemi nj\u00eb p\u00ebrvoj\u00eb t\u00eb gjer\u00eb me sisteme q\u00eb konsumojn\u00eb shum\u00eb memorie, si CAD, EDA dhe t\u00eb ngjashme, t\u00eb cilat kan\u00eb filluar t\u00eb ngadal\u00ebsohen n\u00ebn ngarkesa t\u00eb larta. Dhe ndonj\u00ebher\u00eb kemi hasur probleme me Gluster. Pas nj\u00eb v\u00ebzhgimi t\u00eb kujdessh\u00ebm p\u00ebr disa dit\u00eb mbi memorjen e p\u00ebrdorur dhe koh\u00ebn e pritjes s\u00eb disqeve, ne has\u00ebm n\u00eb ngarkes\u00eb, nj\u00eb iowait shum\u00eb t\u00eb madh, gabime n\u00eb b\u00ebrtham\u00eb (kernel oops), dhe bllokime etj. <\/p>\n

Ky artikull \u00ebsht\u00eb rezultat i eksperimentimeve t\u00eb shumta me konfigurimin e parametrave, t\u00eb kryera n\u00eb situata t\u00eb ndryshme. Fal\u00eb k\u00ebtyre parametrave, jo vet\u00ebm q\u00eb u p\u00ebrmir\u00ebsua p\u00ebrgjigjshm\u00ebria n\u00eb p\u00ebrgjith\u00ebsi, por edhe funksionimi i klasterit u stabilizua ndjesh\u00ebm.<\/p>\n

Kur b\u00ebhet fjal\u00eb p\u00ebr konfigurimin e memories, e para q\u00eb duhet t\u00eb shikoni \u00ebsht\u00eb n\u00ebn-sistemi i memories virtuale (VM, virtual memory), t\u00eb cilin ka shum\u00eb mund\u00ebsi q\u00eb mund t'ju ngat\u00ebrrojn\u00eb.<\/p>\n

vm.swappiness<\/h4>\n

\nParametri vm.swappiness<\/code> p\u00ebrcakton se sa shum\u00eb b\u00ebrthama p\u00ebrdor swap-in (n\u00eb m\u00ebnyr\u00eb krahasuese me memorien operative). N\u00eb kodin burimor, ai gjithashtu \u00ebsht\u00eb p\u00ebrcaktuar si \u00abtendency to steal mapped memory\u00bb (tendenca p\u00ebr t\u00eb vjedhur memorien e mapuar). Nj\u00eb vler\u00eb e lart\u00eb swappiness n\u00ebnkupton se b\u00ebrthama do t\u00eb jet\u00eb m\u00eb e prirur t\u00eb shkarkoj\u00eb faqet e mapuara. Nj\u00eb vler\u00eb e ul\u00ebt swappiness n\u00ebnkupton t\u00eb kund\u00ebrt\u00ebn: b\u00ebrthama do t\u00eb shkarkoj\u00eb m\u00eb pak faqe nga memoria. N\u00eb fjal\u00eb t\u00eb tjera, sa m\u00eb e lart\u00eb t\u00eb jet\u00eb vlera vm.swappiness<\/code>, aq m\u00eb shum\u00eb sistemi do t\u00eb p\u00ebrdor\u00eb swap.<\/p>\n

P\u00ebrdorimi i madh i swap-it \u00ebsht\u00eb i pad\u00ebshiruesh\u00ebm, sepse blloqet e m\u00ebdha t\u00eb t\u00eb dh\u00ebnave ngarkohet dhe shkarkohet nga memoria operative. Shum\u00eb pretendojn\u00eb se vlera e swapiness duhet t\u00eb jet\u00eb e madhe, por nga p\u00ebrvoja ime, instalimi n\u00eb \u00ab0\u00bb \u00e7on n\u00eb nj\u00eb rritje t\u00eb performanc\u00ebs.<\/p>\n

Mund t\u00eb lexoni m\u00eb shum\u00eb k\u00ebtu \u2014 lwn.net\/Articles\/100978<\/a><\/noindex> <\/p>\n

Por, s\u00ebrish, k\u00ebto \u043d\u0430\u0441\u0442\u0440\u043e\u0439\u043a\u0438 duhet t\u00eb aplikohen me kujdes dhe vet\u00ebm pas testimit t\u00eb aplikacionit specifik. P\u00ebr aplikacionet e ngarkuara me flukse t\u00eb larta, ky parameter duhet vendosur n\u00eb \u00ab0\u00bb. Duke e ndryshuar n\u00eb \u00ab0\u00bb, p\u00ebrmir\u00ebsohet p\u00ebrgjigjshm\u00ebria e sistemit.<\/p>\n

vm.vfs_cache_pressure<\/h4>\n

\nKy parameter kontrollon memorien e konsumuar nga b\u00ebrthama p\u00ebr memorizimin e objekteve t\u00eb katalog\u00ebve dhe indeks\u00ebve t\u00eb descriptors (dentry dhe inode).<\/p>\n

Me nj\u00eb vler\u00eb t\u00eb paracaktuar prej 100, b\u00ebrthama do t\u00eb p\u00ebrpiqet t\u00eb \u00e7liroj\u00eb cache-in e dentry dhe inode n\u00eb m\u00ebnyr\u00eb \u00abt\u00eb drejt\u00eb\u00bb n\u00eb lidhje me pagecache dhe swapcache. Ulja e vfs_cache_pressure \u00e7on n\u00eb ruajtje t\u00eb caches dentry dhe inode nga b\u00ebrthama. Kur vlera \u00ebsht\u00eb \u00ab0\u00bb, b\u00ebrthama kurr\u00eb nuk do t\u00eb pastrohet cache-in e dentry dhe inode p\u00ebr shkak t\u00eb munges\u00ebs s\u00eb memories (memory pressure), dhe k\u00ebt\u00eb mund t\u00eb \u00e7oj\u00eb leht\u00ebsisht n\u00eb nj\u00eb gabim out-of-memory. Rritja e vfs_cache_pressure p\u00ebrtej 100 \u00e7on n\u00eb k\u00ebt\u00eb q\u00eb b\u00ebrthama i jep p\u00ebrpar\u00ebsi shkarkimit t\u00eb dentry dhe inode.<\/p>\n

Kur p\u00ebrdoret GlusterFS, shum\u00eb p\u00ebrdorues me sasi t\u00eb m\u00ebdha t\u00eb t\u00eb dh\u00ebnave dhe shum\u00eb skedar\u00eb t\u00eb vegj\u00ebl mund t\u00eb p\u00ebrdorin leht\u00ebsisht nj\u00eb sasi t\u00eb konsiderueshme t\u00eb memories operative n\u00eb server p\u00ebr shkak t\u00eb memorizimit t\u00eb inode\/dentry, q\u00eb mund t\u00eb \u00e7oj\u00eb n\u00eb nj\u00eb r\u00ebnie t\u00eb performanc\u00ebs, pasi b\u00ebrthama duhet t\u00eb procesoj\u00eb strukturat e t\u00eb dh\u00ebnave n\u00eb nj\u00eb sistem me 40 GB memorie. Vendosja e k\u00ebtij parametri mbi 100 ndihmoi shum\u00eb p\u00ebrdorues t\u00eb arrijn\u00eb memorizim m\u00eb t\u00eb drejt\u00eb dhe t\u00eb p\u00ebrmir\u00ebsojn\u00eb p\u00ebrgjigjshm\u00ebrin\u00eb e b\u00ebrtham\u00ebs.<\/p>\n

vm.dirty_background_ratio dhe vm.dirty_ratio<\/h4>\n

\nParametri i par\u00eb (vm.dirty_background_ratio<\/code>) p\u00ebrcakton p\u00ebrqindjen e memories me faqe t\u00eb ndyra, duke arritur t\u00eb cil\u00ebn duhet t\u00eb filloj\u00eb nj\u00eb shkarkim t\u00eb fush\u00ebs s\u00eb ndyshme n\u00eb disk. Deri sa ky p\u00ebrqindje nuk arrihet, faqet nuk shkarkohen n\u00eb disk. Dhe kur fillon shkarkimi, ai realizohet n\u00eb m\u00ebnyr\u00eb t\u00eb prapambetur, pa nd\u00ebrprer\u00eb proceset aktive. <\/p>\n

Parametri i dyt\u00eb (vm.dirty_ratio<\/code>) p\u00ebrcakton p\u00ebrqindjen e memories q\u00eb mund t\u00eb z\u00ebni nga faqet e ndotura deri n\u00eb fillimin e shlyerjes me forc\u00eb (forced flash). Pasi t\u00eb arrihet ky prag, t\u00eb gjith\u00eb proceset b\u00ebhen sinkronike (bllokohen), dhe nuk u lejohet t\u00eb vazhdojn\u00eb pun\u00ebn deri sa operacioni i tyre i k\u00ebrkuar i hyrjes-daljes t\u00eb p\u00ebrfundoj\u00eb ndryshe dhe t\u00eb dh\u00ebnat t\u00eb ndodhin n\u00eb disk. Kur hyni shum\u00eb n\u00eb hyrje-dalje, kjo krijon nj\u00eb problem, pasi ndodhet mungesa e keq\u00ebt\u00eb k\u00ebshillimit t\u00eb t\u00eb dh\u00ebnave dhe t\u00eb gjitha proceset q\u00eb kryejn\u00eb hyn\u00eb-dalje bllokohen duke pritur hyrjen-daljen. Kjo \u00e7on n\u00eb nj\u00eb num\u00ebr t\u00eb madh t\u00eb proceseve t\u00eb bllokuar, ngarkes\u00eb t\u00eb lart\u00eb, funksionim jo stabil t\u00eb sistemit dhe performanc\u00eb t\u00eb dob\u00ebt.<\/p>\n

Ulja e vlerave t\u00eb k\u00ebtyre parametrave \u00e7on n\u00eb faktin q\u00eb t\u00eb dh\u00ebnat p\u00ebrfundojn\u00eb m\u00eb shpesh n\u00eb disk dhe nuk ruhet n\u00eb RAM. Kjo mund t\u2019i ndihmoj\u00eb sistemet me shum\u00eb memorie, p\u00ebr t\u00eb cilat nuk \u00ebsht\u00eb e pazakont\u00eb t\u00eb hiqen p\u00ebr n\u00eb disk cache faqesh prej 45\u201390 GB, gj\u00eb q\u00eb \u00e7on n\u00eb nj\u00eb koh\u00eb t\u00eb madhe prite p\u00ebr aplikacionet frontend, duke reduktuar p\u00ebrgjigjen dhe interaktivitetin e p\u00ebrgjithsh\u00ebm.<\/p>\n

\u00ab1\u00bb > \/proc\/sys\/vm\/pagecache<\/h4>\n

\nCache i faqeve (page cache) \u00ebsht\u00eb cache ku ruhet t\u00eb dh\u00ebnat e fileve dhe programeve t\u00eb ekzekutueshme, pra jan\u00eb faqet me p\u00ebrmbajtje reale t\u00eb fileve ose pajisjeve bllok. Ky cache p\u00ebrdoret p\u00ebr t\u00eb reduktuar numrin e leximeve nga disku. Vlera \u00ab1\u00bb do t\u00eb thot\u00eb se p\u00ebr cache p\u00ebrdoret 1% e RAM dhe operacionet e leximit nga disku do t\u00eb jen\u00eb m\u00eb shum\u00eb sesa nga RAM. Nuk \u00ebsht\u00eb domosdoshm\u00ebri ta ndryshoni k\u00ebt\u00eb parametrash, por n\u00ebse keni nj\u00eb q\u00ebndrim paranoid mbi kontrollimin e cache t\u00eb faqeve, mund ta shfryt\u00ebzoni at\u00eb.<\/p>\n

\u00abdeadline\u00bb > \/sys\/block\/sdc\/queue\/scheduler<\/h4>\n

\nPlanifikuesi i hyrjes-daljes (I\/O scheduler) \u00ebsht\u00eb nj\u00eb komponent i b\u00ebrtham\u00ebs Linux q\u00eb trajton radh\u00ebt p\u00ebr lexim dhe shkrim. N\u00eb teori, p\u00ebr nj\u00eb kontrollues RAID t\u00eb men\u00e7ur \u00ebsht\u00eb m\u00eb mir\u00eb t\u00eb p\u00ebrdoret \u00abnoop\u00bb, sepse Linux nuk di asgj\u00eb p\u00ebr gjeometrin\u00eb fizike t\u00eb disku, andaj \u00ebsht\u00eb m\u00eb efektive t\u2019i lejohet kontrolluesit, q\u00eb e njeh mir\u00eb struktur\u00ebn e disku, t\u00eb p\u00ebrpunoj\u00eb k\u00ebrkesat sa m\u00eb shpejt. Por duket se \u00abdeadline\u00bb e p\u00ebrmir\u00ebson performanc\u00ebn. M\u00eb shum\u00eb p\u00ebr planifikuesit mund t\u00eb lexoni n\u00eb dokumentacionin e kodit burimor t\u00eb b\u00ebrtham\u00ebs Linux: linux\/Documentation\/block\/*osched.txt<\/code>. Dhe gjithashtu kam v\u00ebrejtur nj\u00eb rritje t\u00eb kapacitetit t\u00eb leximit gjat\u00eb operacioneve t\u00eb p\u00ebrziera (shum\u00eb operacione shkrimi).<\/p>\n

\u00ab256\u00bb > \/sys\/block\/sdc\/queue\/nr_requests<\/h4>\n

\nNumri i k\u00ebrkesave t\u00eb input-it dhe output-it n\u00eb buffer para se t'i kalohen planifikuesit. Madh\u00ebsia e radh\u00ebs brenda disa kontrolluesve (queue_depth) \u00ebsht\u00eb m\u00eb e madhe se nr_requests e planifikuesit t\u00eb input-it dhe output-it, k\u00ebshtu q\u00eb planifikuesi ka pak gjasa t\u00eb p\u00ebrprioritoj\u00eb dhe t\u00eb kryej\u00eb korrekt\u00ebsisht bashkimet e k\u00ebrkesave. P\u00ebr planifikuesit deadline dhe CFQ, \u00ebsht\u00eb m\u00eb mir\u00eb kur nr_requests \u00ebsht\u00eb dy her\u00eb m\u00eb shum\u00eb se radh\u00ebs brenda kontrolluesit. Bashkimi dhe riorganizimi i k\u00ebrkesave ndihmon planifikuesin t\u00eb jet\u00eb m\u00eb reagues n\u00eb ngarkesa t\u00eb m\u00ebdha.<\/p>\n

echo \u00ab16\u00bb > \/proc\/sys\/vm\/page-cluster<\/h4>\n

\nParametri page-cluster menaxhon numrin e faqeve q\u00eb shkruhen n\u00eb swap nj\u00ebher\u00ebsh. N\u00eb shembullin e m\u00ebsip\u00ebrm, vlera vendoset n\u00eb \u00ab16\u00bb n\u00eb p\u00ebrputhje me madh\u00ebsin\u00eb e stripe (stripe size) t\u00eb RAID-it q\u00eb \u00ebsht\u00eb 64 KB. Kjo nuk ka kuptim kur swappiness = 0, por n\u00ebse e keni vendosur swappiness n\u00eb 10 ose 20, at\u00ebher\u00eb p\u00ebrdorimi i k\u00ebsaj vlere do t'ju ndihmoj\u00eb kur madh\u00ebsia e stripe-it t\u00eb RAID-it \u00ebsht\u00eb 64 KB.<\/p>\n

blockdev \u2014setra 4096 \/dev\/<<\/code>emri i pajisjes><\/code> (-sdb, hdc ose dev_mapper)<\/h4>\n

\nCaktimet e pajisjeve bllok normale p\u00ebr shum\u00eb kontrollues RAID zakonisht \u00e7ojn\u00eb n\u00eb performanc\u00eb t\u00eb tmerrshme. Shtimi i opsionit t\u00eb m\u00ebsip\u00ebrm konfiguroni leximin e parashikimit p\u00ebr sektor\u00eb prej 4096 * 512 byte. T\u00eb pakt\u00ebn p\u00ebr operacione t\u00eb rrjedhshme, shpejt\u00ebsia rritet p\u00ebrmbush duke mbushur t\u00eb dh\u00ebnat e nd\u00ebrtuara n\u00eb cache-n\u00eb e \u00e7astit duke lexuar n\u00eb m\u00ebnyr\u00eb t\u00eb parashikueshme n\u00eb periudh\u00ebn q\u00eb b\u00ebrthama p\u00ebrdor p\u00ebr p\u00ebrgatitjen e input-it dhe output-it. N\u00eb cache mund t\u00eb ruhen t\u00eb dh\u00ebnat q\u00eb do t\u00eb k\u00ebrkohen gjat\u00eb leximit t\u00eb ardhsh\u00ebm. Leximi i tep\u00ebrt i parashikimit mund t\u00eb vras\u00eb input-in dhe output-in e rast\u00ebsish\u00ebm p\u00ebr skedar\u00eb t\u00eb m\u00ebdhenj, n\u00ebse p\u00ebrdor koh\u00ebn e mundshme t\u00eb diskut ose ngarkon t\u00eb dh\u00ebna p\u00ebrtej cache-it.<\/p>\n

M\u00eb posht\u00eb jan\u00eb disa rekomandime t\u00eb tjera n\u00eb nivelin e sistemit t\u00eb skedar\u00ebve. Por ato ende nuk jan\u00eb testuar. Sigurohuni q\u00eb sistemi juaj i skedar\u00ebve e di madh\u00ebsin\u00eb e stripe-it dhe numrin e disqeve n\u00eb grumbull. P\u00ebr shembull, q\u00eb ky \u00ebsht\u00eb nj\u00eb grumbull raid5 me madh\u00ebsi stripe 64K nga gjasht\u00eb disqe (n\u00eb fakt nga pes\u00eb, sepse nj\u00eb disk p\u00ebrdoret p\u00ebr paritet). K\u00ebto rekomandime jan\u00eb t\u00eb bazuara n\u00eb supozime teorike dhe jan\u00eb mbledhur nga blogje t\u00eb ndryshme\/artikuj ekspert\u00ebsh p\u00ebr RAID.<\/p>\n

-> skedar\u00eb ext4, 5 disqe, 64K stripe, nj\u00ebsit\u00eb n\u00eb blloqe 4K\nmkfs -text4 -E stride=$((64\/4))\n-> xfs, 5 disqe, 64K stripe, nj\u00ebsit\u00eb n\u00eb sektor\u00eb prej 512 byte\nmkfs -txfs -d sunit=$((64*2)) -d swidth=$((5*64*2))<\/code><\/pre>\n
\nP\u00ebr skedar\u00eb t\u00eb m\u00ebdhenj, mund t\u00eb shqyrtohet mund\u00ebsia e rritjes s\u00eb madh\u00ebsive t\u00eb m\u00ebsip\u00ebrme t\u00eb stripeve.<\/p>\n
K\u00cbSHILLIM!<\/b> T\u00eb gjitha q\u00eb jan\u00eb p\u00ebrshkruar m\u00eb sip\u00ebr jan\u00eb jasht\u00ebzakonisht subjektive p\u00ebr disa lloje aplikacionesh. Ky artikull nuk garanton p\u00ebrmir\u00ebsime pa testimin paraprak t\u00eb aplikacioneve p\u00ebrkat\u00ebse nga p\u00ebrdoruesi. Duhet t\u00eb aplikohet vet\u00ebm n\u00eb rast se nevojitet p\u00ebrmir\u00ebsimi i p\u00ebrgjithsh\u00ebm i p\u00ebrgjigjeve t\u00eb sistemit ose n\u00ebse zgjidh probleme aktuale.<\/p>\n
Materiale shtes\u00eb:<\/h4>\n
<\/p>\n