Spark资源调度机制源码分析--基于spreadOutApps及非spreadOutApps两种资源调度算法
Posted 花心土豆
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了Spark资源调度机制源码分析--基于spreadOutApps及非spreadOutApps两种资源调度算法相关的知识,希望对你有一定的参考价值。
Spark资源调度机制源码分析--基于spreadOutApps及非spreadOutApps两种资源调度算法
1、spreadOutApp尽量平均分配到每个executor上;
2、非spreadOutApp尽量在使用单个executor的资源。
源码分析
org.apache.spark.deploy.master.Master
1、首先判断,master状态不是ALIVE的话,直接返回
2、调度driver
3、 Application的调度机制(核心之核心,重中之重)
源码如下:
1 /* 2 *schedule()解决了spark资源调度的问题 3 */ 4 private def schedule() { 5 //首先判断,master状态不是ALIVE的话,直接返回 6 //也就是说,stanby master是不会进行application等资源调度的 7 if (state != RecoveryState.ALIVE) { return } 8 9 // First schedule drivers, they take strict precedence over applications 10 // Randomization helps balance drivers 11 12 //Random.shuffle的原理,大家要清楚,就是对传入的集合的元素进行随机的打乱 13 //取出了workers中的所有之前注册上来的worker,进行过滤,必须是状态为ALIVE的worker 14 //对状态为ALIVE的worker,调用Random的shuffle方法进行随机的打乱 15 val shuffledAliveWorkers = Random.shuffle(workers.toSeq.filter(_.state == WorkerState.ALIVE)) 16 val numWorkersAlive = shuffledAliveWorkers.size 17 var curPos = 0 18 19 //首先,调度driver 20 //为什么要调度driver,大家想一下,什么情况下,会注册driver,并且会导致driver被调度 21 //其实 ,只有用yarn-cluster模式提交的时候,才会注册driver;因为standalone和yarn-client模式,都会在本地直接 22 //启动driver,而不会来注册driver,就更不可能让master调度driver了 23 24 //driver调度机制 25 //遍历waittingDrivers ArrayBuffer 26 for (driver <- waitingDrivers.toList) { // iterate over a copy of waitingDrivers 27 // We assign workers to each waiting driver in a round-robin fashion. For each driver, we 28 // start from the last worker that was assigned a driver, and continue onwards until we have 29 // explored all alive workers. 30 var launched = false 31 var numWorkersVisited = 0 32 33 //while的条件,numWorkersVisited小于numWorkersAlive 34 //什么意思?就是说,只要还有活着的worker没有遍历到,那么就继续进行遍历 35 //而且,当前这个driver还没有被启动,也就是launched为false 36 while (numWorkersVisited < numWorkersAlive && !launched) { 37 val worker = shuffledAliveWorkers(curPos) 38 numWorkersVisited += 1 39 40 //如果当前这个worker的空闲内存量大于等于,driver需要的内存 41 //并且worker的空闲cpu数量,大于等于driver需要的cpu数量 42 if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) { 43 //启动driver 44 launchDriver(worker, driver) 45 //并且将driver从waitingDrivers队列中移除 46 waitingDrivers -= driver 47 launched = true 48 } 49 50 //将指针指向下一个worker 51 curPos = (curPos + 1) % numWorkersAlive 52 } 53 } 54 55 // Right now this is a very simple FIFO scheduler. We keep trying to fit in the first app 56 // in the queue, then the second app, etc. 57 // Application的调度机制(核心之核心,重中之重) 58 // 首先, application的调度算法有两种,一种是spreadOutApps,另一种是非spreadOutApps 59 if (spreadOutApps) { 60 // Try to spread out each app among all the nodes, until it has all its cores 61 62 //首先,遍历waitingApps中的ApplicationInfo,并且过滤出application还需要高度的cores的application 63 for (app <- waitingApps if app.coresLeft > 0) { 64 //从workers中,过滤状态为ALIVE的,再次过滤可以被Application使用的Worker,然后按照剩余cpu数量倒序排序 65 val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE) 66 .filter(canUse(app, _)).sortBy(_.coresFree).reverse 67 val numUsable = usableWorkers.length 68 //创建一个空数组,存储了要分配给每个worker的cpu数量 69 val assigned = new Array[Int](numUsable) // Number of cores to give on each node 70 //获取到底要分配多少cpu,取app剩余要分配的cpu的数量和worker总共可用cpu数量的最小值 71 var toAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum) 72 73 //通过这种算法,其实会将每个application,要启动的executor,都平均分布到各个worker上去 74 //比如有20个cpu core要分配,那么实际会循环两遍worker,每次循环,给每个worker分配1个core 75 //最后每个worker分配了2个core 76 77 //while条件,只要要分配的cpu,还没有分配完,就继续循环 78 var pos = 0 79 while (toAssign > 0) { 80 //每一个worker,如果空闲的cpu数量大于,已经分配出去的cpu数量 81 //也就是说,worker还有可分配的cpu 82 if (usableWorkers(pos).coresFree - assigned(pos) > 0) { 83 //将总共要分配的cpu数量-1,因为这里已经决定在这个worker上分配一个cpu了 84 toAssign -= 1 85 //给这个worker分配的cpu数量,加1 86 assigned(pos) += 1 87 } 88 //指针移动到下一下worker 89 pos = (pos + 1) % numUsable 90 } 91 92 // Now that we‘ve decided how many cores to give on each node, let‘s actually give them 93 // 给每个worker分配完application要求的cpu core之后 94 // 遍历worker 95 for (pos <- 0 until numUsable) { 96 //只要判断之前给这个worker分配到了core 97 if (assigned(pos) > 0) { 98 //首先,在application内部缓存结构中,添加executor 99 //并且创建ExecutorDesc对象,其中封装了,给这个executor分配多少个cpu core 100 //在spark-submit脚本中,可以指定要多少executor,每个execuor多少个cpu,多少内存 101 //那么基于源码机制,实际上,executor的实际数量,以及每个executor的cpu,可能与配置是不一样的 102 //因为,我人帝里基于总的cpu来分配的,就是比如,要求3个executor,每个要3个cpu,那么比如,有9个workers,每个有1个cpu 103 //那么其实总其知道,要分配9个core,其实根据这种算法,会给每个worker分配一个core,然后给每个worker启动一个executor 104 //最后会启动,9个executor,每个executor有1个cpu core 105 val exec = app.addExecutor(usableWorkers(pos), assigned(pos)) 106 //那么就在worker上启动executor 107 launchExecutor(usableWorkers(pos), exec) 108 //将application状态设置为running 109 app.state = ApplicationState.RUNNING 110 } 111 } 112 } 113 } else { 114 // Pack each app into as few nodes as possible until we‘ve assigned all its cores 115 116 //非spreadOutApps调度算法 117 118 //这种算法与spreadOutApps算法正好相反,1、spreadOutApp尽量平均分配到每个executor上;2、非spreadOutApp尽量在使用单个executor的资源。 119 //每个application,都尽可能分配到尽量少的worker上去,比如总其有10个worker,每个有10个core 120 //app总共要分配 20个core,那么其实,只会分配到两个worker上,每个worker都占满10个core 121 //那么,其余的app,就只能 分配到下一个worker了 122 //比如,spark-submit里,配置的是要10个executor,每个要2个core,那么总共是20个croe 123 //只会启动2个executor,每个有10个cores 124 125 //将每个Application,尽可能少的分配到worker上去 126 //首先,遍历worker,并且是状态为ALIVE,还有空闲cpu的worker 127 for (worker <- workers if worker.coresFree > 0 && worker.state == WorkerState.ALIVE) { 128 //遍历application,并且是还有城朵分配的core的application 129 for (app <- waitingApps if app.coresLeft > 0) { 130 //判断,如果当前这个worker可以被 application使用 131 if (canUse(app, worker)) { 132 //取worker剩余cpu数量,与app要分配的cpu数量的最小值 133 val coresToUse = math.min(worker.coresFree, app.coresLeft) 134 //如果Worker剩余cpu为0了,就不分配了 135 if (coresToUse > 0) { 136 // 给app添加一个executor 137 val exec = app.addExecutor(worker, coresToUse) 138 //在worker上启动executor 139 launchExecutor(worker, exec) 140 //将application状态设置为running 141 app.state = ApplicationState.RUNNING 142 } 143 } 144 } 145 } 146 } 147 }
以上是关于Spark资源调度机制源码分析--基于spreadOutApps及非spreadOutApps两种资源调度算法的主要内容,如果未能解决你的问题,请参考以下文章
第31课:Spark资源调度分配内幕天机彻底解密:Driver在Cluster模式下的启动两种不同的资源调度方式源码彻底解析资源调度内幕总结