本套系列博客從真實(shí)商業(yè)環(huán)境抽取案例進(jìn)行總結(jié)和分享,并給出Spark商業(yè)應(yīng)用的調(diào)優(yōu)建議和集群環(huán)境容量規(guī)劃等內(nèi)容,請(qǐng)持續(xù)關(guān)注本套博客。版權(quán)聲明:本套Spark調(diào)優(yōu)系列版權(quán)歸作者(秦凱新)所有,禁止轉(zhuǎn)載,歡迎學(xué)習(xí)。
Spark商業(yè)環(huán)境實(shí)戰(zhàn)及調(diào)優(yōu)進(jìn)階系列
- Spark商業(yè)環(huán)境實(shí)戰(zhàn)-Spark內(nèi)置框架rpc通訊機(jī)制及RpcEnv基礎(chǔ)設(shè)施
- Spark商業(yè)環(huán)境實(shí)戰(zhàn)-Spark事件監(jiān)聽(tīng)總線流程分析
- Spark商業(yè)環(huán)境實(shí)戰(zhàn)-Spark存儲(chǔ)體系底層架構(gòu)剖析
- Spark商業(yè)環(huán)境實(shí)戰(zhàn)-Spark底層多個(gè)MessageLoop循環(huán)線程執(zhí)行流程分析
1. Spark存儲(chǔ)體系組件關(guān)系解釋
BlockInfoManger 主要提供讀寫(xiě)鎖控制,層級(jí)僅僅位于BlockManger之下,通常Spark讀寫(xiě)操作都先調(diào)用BlockManger,然后咨詢BlockInfoManger是否存在鎖競(jìng)爭(zhēng),然后才會(huì)調(diào)用DiskStore和MemStore,進(jìn)而調(diào)用DiskBlockManger來(lái)確定數(shù)據(jù)與位置映射,或者調(diào)用 MemoryManger來(lái)確定內(nèi)存池的軟邊界和內(nèi)存使用申請(qǐng)。
1.1 Driver 與 Executor 與 SparkEnv 與 BlockManger 組件關(guān)系:
Driver與 Executor 組件各自擁有任務(wù)執(zhí)行的SparkEnv環(huán)境,而每一個(gè)SparkEnv 中都有一個(gè)BlockManger負(fù)責(zé)存儲(chǔ)服務(wù),作為高層抽象,BlockManger 之間需要通過(guò) RPCEnv,ShuffleClient,及BlocakTransferService相互通訊。
1.1 BlockInfoManger 與 BlockInfo 共享鎖和排它鎖讀寫(xiě)控制關(guān)系:
BlockInfo中具有讀寫(xiě)鎖的標(biāo)志,通過(guò)標(biāo)志可以判斷是否進(jìn)行寫(xiě)控制
val NO_WRITER: Long = -1
val NON_TASK_WRITER: Long = -1024
* The task attempt id of the task which currently holds the write lock for this block, or
* [[BlockInfo.NON_TASK_WRITER]] if the write lock is held by non-task code, or
* [[BlockInfo.NO_WRITER]] if this block is not locked for writing.
def writerTask: Long = _writerTask
def writerTask_=(t: Long): Unit = {
_writerTask = t
checkInvariants()
BlockInfoManager具有BlockId與BlockInfo的映射關(guān)系以及任務(wù)id與BlockId的鎖映射:
private[this] val infos = new mutable.HashMap[BlockId, BlockInfo]
*Tracks the set of blocks that each task has locked for writing.
private[this] val writeLocksByTask = new mutable.HashMap[TaskAttemptId, mutable.Set[BlockId]]
with mutable.MultiMap[TaskAttemptId, BlockId]
*Tracks the set of blocks that each task has locked for reading, along with the number of times
*that a block has been locked (since our read locks are re-entrant).
private[this] val readLocksByTask =
new mutable.HashMap[TaskAttemptId, ConcurrentHashMultiset[BlockId]]
1.3 DiskBlockManager 與 DiskStore 組件關(guān)系:
可以看到DiskStore內(nèi)部會(huì)調(diào)用DiskBlockManager來(lái)確定Block的讀寫(xiě)位置:
-
以下是DiskStore的抽象寫(xiě)操作,需要傳入FileOutputStream => Unit高階函數(shù):
def put(blockId: BlockId)(writeFunc: FileOutputStream => Unit): Unit = { if (contains(blockId)) { throw new IllegalStateException(s"Block $blockId is already present in the disk store") } logDebug(s"Attempting to put block $blockId") val startTime = System.currentTimeMillis val file = diskManager.getFile(blockId) val fileOutputStream = new FileOutputStream(file) var threwException: Boolean = true try { writeFunc(fileOutputStream) threwException = false } finally { try { Closeables.close(fileOutputStream, threwException) } finally { if (threwException) { remove(blockId) } } } val finishTime = System.currentTimeMillis logDebug("Block %s stored as %s file on disk in %d ms".format( file.getName, Utils.bytesToString(file.length()), finishTime - startTime)) } -
以下是DiskStore的讀操作,調(diào)用DiskBlockManager來(lái)獲取數(shù)據(jù)位置:
def getBytes(blockId: BlockId): ChunkedByteBuffer = { val file = diskManager.getFile(blockId.name) val channel = new RandomAccessFile(file, "r").getChannel Utils.tryWithSafeFinally { * For small files, directly read rather than memory map if (file.length < minMemoryMapBytes) { val buf = ByteBuffer.allocate(file.length.toInt) channel.position(0) while (buf.remaining() != 0) { if (channel.read(buf) == -1) { throw new IOException("Reached EOF before filling buffer\n" + s"offset=0\nfile=${file.getAbsolutePath}\nbuf.remaining=${buf.remaining}") } } buf.flip() new ChunkedByteBuffer(buf) } else { new ChunkedByteBuffer(channel.map(MapMode.READ_ONLY, 0, file.length)) } } { channel.close() } }
1.3 MemManager 與 MemStore 與 MemoryPool 組件關(guān)系:
在這里要強(qiáng)調(diào)的是:第一代大數(shù)據(jù)框架hadoop只將內(nèi)存作為計(jì)算資源,而Spark不僅將內(nèi)存作為計(jì)算資源外,還將內(nèi)存的一部分納入存儲(chǔ)體系:
- 內(nèi)存池模型 :邏輯上分為堆內(nèi)存和堆外內(nèi)存,然后堆內(nèi)存(或堆外內(nèi)存)內(nèi)部又分為StorageMemoryPool和ExecutionMemoryPool。
- MemManager是抽象的,定義了內(nèi)存管理器的接口規(guī)范,方便以后擴(kuò)展,比如:老版的StaticMemoryManager和新版的UnifiedMemoryManager.
- MemStore 依賴于UnifiedMemoryManager進(jìn)行內(nèi)存的申請(qǐng)和軟邊界變化或內(nèi)存釋放。
- MemStore 內(nèi)部同時(shí)負(fù)責(zé)存儲(chǔ)真實(shí)的對(duì)象,比如內(nèi)部成員變量:entries ,建立了內(nèi)存中的BlockId與MemoryEntry(Block的內(nèi)存的形式)之間的映射。
- MemStore 內(nèi)部的“占座”行為,如:內(nèi)部變量offHeapUnrollMemoryMap 和onHeapUnrollMemoryMap。
1.4 BlockManagerMaster 與 BlockManager 組件關(guān)系:
- BlockManagerMaster的作用就是對(duì)存在于Dirver或Executor上的BlockManger進(jìn)行統(tǒng)一管理,這簡(jiǎn)直是代理行為,因?yàn)樗钟蠦lockManagerMasterEndpointREf,進(jìn)而和BlockManagerMasterEndpoint進(jìn)行通訊。
2. Spark存儲(chǔ)體系組件BlockTransferServic傳輸服務(wù)
未完待續(xù)
3. 總結(jié)
存儲(chǔ)體系是Spark的基石,我爭(zhēng)取把每一塊細(xì)微的知識(shí)點(diǎn)進(jìn)行剖析,和大部分博客不同的是,我會(huì)盡量采用最平實(shí)的語(yǔ)言,畢竟技術(shù)就是一層窗戶紙。
秦凱新 20181031 凌晨