|
關(guān)于我們
 書單推薦 新書推薦 |
大話存儲后傳/次世代數(shù)據(jù)存儲思維與技術(shù) 
定 價：89 元
9 7 4 8 6 7 4 3 9 0 2 2 1 
全書分為: 靈活的數(shù)據(jù)布局、應用感知及可視化存儲智能、存儲類芯片、儲海鉤沉、集群和多控制器、傳統(tǒng)存儲系統(tǒng)、新興存儲系統(tǒng)、大話光存儲系統(tǒng)、體系結(jié)構(gòu)、IO協(xié)議棧及性能分析、存儲軟件、固態(tài)存儲幾個大章, 其中每章又有多個小節(jié)。每一個小節(jié)都是一個獨立的課題。
冬瓜哥對技術(shù)的追求已經(jīng)到了“癡迷”的境界，與10年前相比，文筆解析更為到位，技術(shù)理解更為精準。其公眾號的每篇文章，都是存儲業(yè)界風向標。
冬瓜哥（張冬），現(xiàn)任某半導體公司系統(tǒng)架構(gòu)師，著有《大話存儲》系列圖書。存儲領域技術(shù)專家和布道者。
第一章 靈活的數(shù)據(jù)布局 ·········································································1
1.1 Raid1.0和Raid1.5 ······························································································2 1.2 Raid5EE和Raid2.0 ·····························································································4 1.3 Lun2.0/SmartMotion ························································································13 第二章 應用感知及可視化存儲智能 ·····················································23 2.1 應用感知精細化自動存儲分層······································································25 2.2 應用感知精細化SmartMotion ········································································27 2.3 應用感知精細化QoS ······················································································28 2.4 產(chǎn)品化及可視化展現(xiàn)······················································································31 2.5 包裝概念制作PPT ···························································································43 2.6 評浪潮“活性”存儲概念··············································································49 第三章 存儲類芯片 ··············································································53 3.1 通道及Raid控制器架構(gòu) ··················································································54 3.2 SAS Expander架構(gòu) ··························································································60 第四章 儲海鉤沉 ··················································································65 4.1 你絕對想不到的兩種高格調(diào)存儲器······························································66 4.2 JBOD里都有什么····························································································70 4.3 Raid4校驗盤之殤 ····························································································72 4.4 為什么說Raid卡是臺小電腦 ··········································································73 4.5 為什么Raid卡電池被換為超級電容 ······························································74 4.6 固件和微碼到底什么區(qū)別··············································································75 4.7 FC成環(huán)器內(nèi)部真的是個環(huán)嗎 ·········································································76 4.8 為什么說SAS、FC對CPU耗費比TCPIP+以太網(wǎng)低 ····································77 4.9 雙控存儲之間的心跳線都跑了哪些流量······················································78 第五章集群和多控制器 ······································································· 79 5.1 淺談雙活和多路徑··························································································80 5.2 “淺”談容災和雙活數(shù)據(jù)中心（上）··························································82 5.3 “淺”談容災和雙活數(shù)據(jù)中心（下）··························································87 5.4 集群文件系統(tǒng)架構(gòu)演變深度梳理圖解··························································96 5.5 從多控緩存管理到集群鎖············································································107 5.6 共享式與分布式各論····················································································115 5.7 “冬瓜哥畫PPT”雙活是個坑 ·····································································118 第六章傳統(tǒng)存儲系統(tǒng) ········································································· 121 6.1 與存儲系統(tǒng)相關(guān)的一些基本話題分享························································122 6.2 高端存儲系統(tǒng)江湖風云錄！········································································133 6.3 驚了！原來高端存儲架構(gòu)是這樣演進的！················································145 6.4 傳統(tǒng)高端存儲系統(tǒng)把數(shù)據(jù)緩存集中外置一石三鳥····································155 6.5 傳統(tǒng)外置存儲已近黃昏················································································156 6.6 存儲圈老炮大戰(zhàn)小鮮肉················································································166 6.7 傳統(tǒng)存儲老矣，新興存儲能當大任否？····················································167 第七章次世代存儲系統(tǒng) ····································································· 185 7.1 一桿老槍照玩次世代存儲系統(tǒng)····································································187 7.2 最有傳統(tǒng)存儲格調(diào)的次世代存儲系統(tǒng)························································192 7.3 最適合大規(guī)模數(shù)據(jù)中心的次世代存儲系統(tǒng)················································203 7.4 最高性能的次世代存儲系統(tǒng)········································································206 7.5 最具備感知應用能力的次世代存儲系統(tǒng)····················································214 7.6 最具有數(shù)據(jù)管理靈活性的次時代存儲系統(tǒng)················································225 第八章光存儲系統(tǒng)············································································ 237 8.1 光存儲基本原理····························································································238 8.2 神秘的激光頭及藍光技術(shù)············································································244 8.3 剖析藍光存儲系統(tǒng)························································································249 8.4 光存儲系統(tǒng)生態(tài)····························································································253 8.5 站在未來看現(xiàn)在····························································································259 第九章體系結(jié)構(gòu) ················································································ 263 9.1 大話眾核心處理器體系結(jié)構(gòu)········································································264 9.2 致敬龍芯！冬瓜哥手工設計了一個CPU譯碼器！ ····································271 9.3 NUNA體系結(jié)構(gòu)首次落地InCloudRack機柜 ···············································274 9.4 評宏杉科技的CloudSAN架構(gòu) ······································································278 9.5 內(nèi)存竟然還能這么玩？！············································································283 9.6 PCIe交換，什么鬼？····················································································293 9.7 聊聊FPGA/GPCPU/PCIe/Cache-Coherency ················································300 9.8 【科普】超算到底是怎樣算的？································································305 第十章 I/O 協(xié)議棧及性能分析 ···························································· 317 10.1 最完整的存儲系統(tǒng)接口/協(xié)議/連接方式總結(jié) ···········································318 10.2 I/O協(xié)議棧前沿技術(shù)研究動態(tài) ····································································332 10.3 Raid組的Stripe Size到底設置為多少合適？ ·············································344 10.4 并發(fā)I/O——系統(tǒng)性能的根本！ ································································347 10.5 關(guān)于I/O時延你被騙了多久？ ····································································349 10.6 如何測得整條I/O路徑上的并發(fā)度？ ························································351 10.7 隊列深度、時延、并發(fā)度、吞吐量的關(guān)系到底是什么··························351 10.8 為什么Raid對于某些場景沒有任何提速作用？ ······································365 10.9 為什么測試時性能出色，上線時卻慘不忍睹？······································366 10.10 隊列深度過淺有什么影響？····································································368 10.11 隊列深度調(diào)節(jié)為多大最理想？ ································································369 10.12 機械盤的隨機I/O平均時延為什么有一過性降低？ ······························370 10.13 數(shù)據(jù)布局到底是怎么影響性能的？························································371 10.14 關(guān)于同步I/O與阻塞I/O的誤解 ·································································374 10.15 原子寫，什么鬼？！················································································375 10.16 何不做個USB Target？ ·············································································385 10.17 冬瓜哥的一項新存儲技術(shù)專利已正式通過············································385 10.18 小梳理一下iSCSI底層 ··············································································394 10.19 FC的4次Login過程簡析 ···········································································396 第十一章存儲軟件············································································ 397 11.1 Thin就是個坑誰用誰找抽！······································································398 11.2 存儲系統(tǒng)OS變遷 ·························································································400 第十二章固態(tài)存儲············································································ 409 12.1 淺析固態(tài)介質(zhì)在存儲系統(tǒng)中的應用方式··················································410 12.2 關(guān)于SSD元數(shù)據(jù)及掉電保護的誤解··························································420 12.3 關(guān)于閃存FTL的Host Base和Device Based的誤解 ····································421 12.4 關(guān)于SSD HMB與CMB ···············································································423 12.5 同有科技展翅歸來······················································································424 12.6 和老唐說相聲之SSD性能測試之“玉”··················································435 12.7 固態(tài)盤到底該怎么做Raid？ ······································································441 12.8 當Raid2.0遇上全固態(tài)存儲 ·········································································448 12.9 上/下頁、快/慢頁、MSB/LSB都些什么鬼？ ··········································451 12.10 關(guān)于對MSB/LSB寫0時的步驟 ·································································457
1.1 Raid1.0和Raid1.5
在機械盤時代，影響最終I/O性能的根本因素無非就是兩個，一個是頂端源頭，也就是應用的I/O調(diào)用方式和I/O屬性；另一個是底端源頭，那就是數(shù)據(jù)最終是以什么形式、狀態(tài)存放在多少機械盤上的。應用如何I/O調(diào)用完全不是存儲系統(tǒng)可以控制的事情，所以從這個源頭來解決性能問題對于存儲系統(tǒng)來講是無法做什么工作的。但是數(shù)據(jù)如何組織、排布，絕對是存儲系統(tǒng)重中之重的工作。 這一點從Raid誕生開始就一直在不斷的演化當中。舉個最簡單的例子，從Raid3到Raid4再到Raid5，Raid3當時設計的時候致力于單線程大塊連續(xù)地址I/O吞吐量最大化，為了實現(xiàn)這個目的，Raid3的條帶非常窄，窄到每次上層下發(fā)的I/O目標地址基本上都落在了所有盤上，這樣幾乎每個I/O都會讓多個盤并行讀寫來服務于這個I/O，而其他I/O就必須等待，所以我們說Raid3陣列場景下，上層的I/O之間是不能并發(fā)的，但是單個I/O是可以采用多盤為其并發(fā)的。所以，如果系統(tǒng)內(nèi)只有一個線程（或者說用戶、程序、業(yè)務），而且這個線程是大塊連續(xù)地址I/O追求吞吐量的業(yè)務，那么Raid3非常合適。但是大部分業(yè)務其實不是這樣，而是追求上層的I/O能夠充分地并行執(zhí)行，比如多線程、多用戶發(fā)出的I/O能夠并發(fā)地被響應，此時就需要增大條帶到一個合適的值，讓一個I/O目標地址范圍不至于牽動Raid組中所有盤為其服務，這樣就有一定幾率讓一組盤同時響應多個I/O，而且盤數(shù)越多，并發(fā)幾率就越大。Raid4相當于條帶可調(diào)的Raid3，但是Raid4獨立校驗盤的存在不但讓其成為高故障率的熱點盤，而且也制約了本可以并發(fā)的I/O，因為伴隨著每個I/O的執(zhí)行，校驗盤上對應條帶的校驗塊都需要被更新，而由于所有校驗塊只存放在這塊盤上，所以上層的I/O只能一個一個第一章 靈活的數(shù)據(jù)布局3地順著執(zhí)行，不能并發(fā)。Raid5則通過把校驗塊打散在Raid組中所有磁盤上，從而實現(xiàn)了并發(fā)I/O。大部分存儲廠商提供針對條帶寬度的設置，比如從32KB到128KB。假設一個I/O請求讀16KB，在一個8塊盤做的Raid5組里，如果條帶為32KB，則每塊盤上的段（Segment）為4KB，這個I/O起碼要占用4塊盤，假設并發(fā)幾率為100%，那么這個Raid組能并發(fā)兩個16KB的I/O，并發(fā)8個4KB的I/O；如果將條帶寬度調(diào)節(jié)為128KB，則在100%并發(fā)幾率的條件下可并發(fā)8個小于等于16KB的I/O。 講到這里，我們可以看到單單是調(diào)節(jié)條帶寬度，以及優(yōu)化校驗塊的布局，就可以得到迥異的性能表現(xiàn)。但是再怎么折騰，I/O性能始終受限在Raid組那少得可憐的幾塊或者十幾塊盤上。為什么是幾塊或者十幾塊？難道不能把100塊盤做成一個大Raid5組，然后，通過把所有邏輯卷創(chuàng)建在它上面來增加每個邏輯卷的性能么？你不會選擇這么做的，當一旦有一塊盤壞掉，系統(tǒng)需要重構(gòu)的時候，你會后悔當時的決定，因為你會發(fā)現(xiàn)此時整個系統(tǒng)性能大幅降低，哪個邏輯卷也別想好過，因為此時99塊盤都 在全速讀出數(shù)據(jù)，系統(tǒng)計算xor校驗塊，然后把校驗塊寫入熱備盤中。當然，你可以控制降速重構(gòu)，來緩解在線業(yè)務的I/O性能，但是付出的代價就是增加了重構(gòu)時間，重構(gòu)周期內(nèi)如果有盤再壞，那么全部數(shù)據(jù)蕩然無存。所以，必須縮小故障影響域，所以一個Raid組最好是幾塊或者十幾塊盤。這比較尷尬，所以人們想出了解決辦法，那就是把多個小Raid5/6組拼接成大Raid0，也就是Raid50/60，然后將邏輯卷分布在其上。當然，目前的存儲廠商黔驢技窮，再也弄出什么新花樣，所以它們習慣把這個大Raid50/60組成“Pool”，也就是池，從而迷惑一部分人，認為存儲又在革新了，存儲依然生命力旺盛。 那冬瓜哥在這里也不妨順水推舟忽悠一下，如果把傳統(tǒng)的Raid組叫作Raid1.0，把Raid50/60叫作Raid1.5。我們其實在這里可以體會出一種周期式上升的規(guī)律，早期盤數(shù)較少，主要靠條帶寬度來調(diào)節(jié)不同場景的性能；后來人們想通了，為何不用Raid50呢？ 把數(shù)據(jù)直接分布到幾百塊盤中，豈不快哉？上層的并發(fā)線程I/O在底層可以實現(xiàn)大規(guī)模并發(fā)，達到超高吞吐量。此時，人們被成功沖昏了頭腦，沒人再去考慮另一個可怕的問題。至這些文字傾諸筆端時仍沒有人考慮這個問題，至少從廠商的產(chǎn)品動向里沒有看出。究其原因，可能是另一輪底層的演變，那就是固態(tài)介質(zhì)。底層的車輪是不斷地提速的，上層的形態(tài)是循環(huán)往復的，但有時候上層可能直接跨越式前進，跨越了其中應該有的一個形態(tài)，這個形態(tài)或者轉(zhuǎn)瞬即逝，亦或者根本沒出現(xiàn)過，但是總會有人產(chǎn)生火花，即便這火花是那么微弱。這個可怕的問題其實被一個更可怕的問題蓋過了，這個更可怕的問題就是重構(gòu)時間過長。一塊4TB的SATA盤，在重構(gòu)的時候就算全速寫入，其轉(zhuǎn)速決定了其吞吐量極4 大話存儲后傳——次世代數(shù)據(jù)存儲思維與技術(shù)限也基本在80MB/s左右，可以算一下，需要58h，實際中為了保證在線業(yè)務的性能，一般會限制在中速重構(gòu)，也就是40MB/s左右，此時需要116h，也就是5天5夜，我敢打賭沒有哪個系統(tǒng)管理員能在這一周內(nèi)睡好覺。 1.2 Raid5EE和Raid2.0 20年前有人發(fā)明過一種叫作Raid5EE的技術(shù)，其目的有兩個，第一是把平時閑著沒事干的熱備盤用起來，第二就是加速重構(gòu)。很顯然，如果把下圖中用“H（hot spare）”表示的熱備盤的空間也像校驗盤一樣，打散到所有盤上的話，就會變成圖右側(cè)所示的布局，每個P塊都跟著一個H塊。這樣整個Raid組能比原來多一塊磁盤可用于工作。另外，由于H空間也被打散了，當有一塊盤損壞時，重構(gòu)的速度理應被加快，因為此時可以多盤并發(fā)寫入了。但是實際卻不然，整個系統(tǒng)的重構(gòu)速度其實并不是被這塊單獨的熱備盤限制了，而是被所有盤一起限制了，因為熱備盤以滿速率寫入重構(gòu)后的數(shù)據(jù)的前提是，其他所有盤都以滿速率讀出數(shù)據(jù)，然后系統(tǒng)對其做xor。就算把熱備盤打散，甚至把熱備盤換成SSD、內(nèi)存，對結(jié)果也毫無影響。那到底怎樣才能加速重構(gòu)呢？唯一的辦法只有像下圖所示這樣，把原本擠在5塊盤里的條帶，橫向打散，請注意，是以條帶為粒度打散，打散單盤是毫無用處的。這樣，才能成倍地提升重構(gòu)速度。
你還可能感興趣
我要評論
|
