Storm自定义调度器实现--DirectScheduler
前言
最近在研究Storm的任务调度相关的知识，于是就想要试着去改造一下Storm的任务调度，来满足一下现实状况中的一些场景。
Storm调度的相关术语
在看Storm的Scheduler代码么之前，得要弄明白几个概念，这样可以帮助大家更好的理解后面的调度过程。
1、slot。这代表一个Supervisor节点上的一个单位资源。每个slot对应一个port，一个slot只能被一个Worker占用。
2、Worker，Executor.Task,1个Worker包含1个或多个Executor执行器，每个执行器包含多个Task。
3、Executor的表现形式为[1-1],[2-2],中括号内的数字代表该Executor中的起始Task id到末尾Task id,1个Worker就相当于在外面加个大括号{[1-1],[2-2]}
4.Component。Storm中的每个组件就是指一类Spout或1个类型的Bolt，这里指的是名称类型，不包含个数。
下面是调度器的核心实现。
代码实现
import backtype.storm.scheduler.*;
import clojure.lang.PersistentArrayMap;
import java.util.*;
/**
* 直接分配调度器,可以分配组件到指定节点中
* Created by zhexuan on 15/7/6.
*/
public class DirectScheduler implements IScheduler{
@Override
public void prepare(Map conf) {
}
@Override
public void schedule(Topologies topologies, Cluster cluster) {
System.out.println("DirectScheduler: begin scheduling");
// Gets the topology which we want to schedule
Collection<TopologyDetails> topologyDetailes;
TopologyDetails topology;
//作业是否要指定分配的标识
String assignedFlag;
Map map;
Iterator<String> iterator = null;
topologyDetailes = topologies.getTopologies();
for(TopologyDetails td: topologyDetailes){
map = td.getConf();
assignedFlag = (String)map.get("assigned_flag");
//如何找到的拓扑逻辑的分配标为1则代表是要分配的,否则走系统的调度
if(assignedFlag != null && assignedFlag.equals("1")){
System.out.println("finding topology named " + td.getName());
topologyAssign(cluster, td, map);
}else {
System.out.println("topology assigned is null");
}
}
//其余的任务由系统自带的调度器执行
new EvenScheduler().schedule(topologies, cluster);
}
/**
* 拓扑逻辑的调度
* @param cluster
* 集群
* @param topology
* 具体要调度的拓扑逻辑
* @param map
* map配置项
*/
private void topologyAssign(Cluster cluster, TopologyDetails topology, Map map){
Set<String> keys;
PersistentArrayMap designMap;
Iterator<String> iterator;
iterator = null;
// make sure the special topology is submitted,
if (topology != null) {
designMap = (PersistentArrayMap)map.get("design_map");
if(designMap != null){
System.out.println("design map size is " + designMap.size());
keys = designMap.keySet();
iterator = keys.iterator();
System.out.println("keys size is " + keys.size());
}
if(designMap == null || designMap.size() == 0){
System.out.println("design map is null");
}
boolean needsScheduling = cluster.needsScheduling(topology);
if (!needsScheduling) {
System.out.println("Our special topology does not need scheduling.");
} else {
System.out.println("Our special topology needs scheduling.");
// find out all the needs-scheduling components of this topology
Map<String, List<ExecutorDetails>> componentToExecutors = cluster.getNeedsSchedulingComponentToExecutors(topology);
System.out.println("needs scheduling(component->executor): " + componentToExecutors);
System.out.println("needs scheduling(executor->components): " + cluster.getNeedsSchedulingExecutorToComponents(topology));
SchedulerAssignment currentAssignment = cluster.getAssignmentById(topology.getId());
if (currentAssignment != null) {
System.out.println("current assignments: " + currentAssignment.getExecutorToSlot());
} else {
System.out.println("current assignments: {}");
}
String componentName;
String nodeName;
if(designMap != null && iterator != null){
while (iterator.hasNext()){
componentName = iterator.next();
nodeName = (String)designMap.get(componentName);
System.out.println("现在进行调度 组件名称->节点名称:" + componentName + "->" + nodeName);
componentAssign(cluster, topology, componentToExecutors, componentName, nodeName);
}
}
}
}
}
/**
* 组件调度
* @param cluster
* 集群的信息
* @param topology
* 待调度的拓扑细节信息
* @param totalExecutors
* 组件的执行器
* @param componentName
* 组件的名称
* @param supervisorName
* 节点的名称
*/
private void componentAssign(Cluster cluster, TopologyDetails topology, Map<String, List<ExecutorDetails>> totalExecutors, String componentName, String supervisorName){
if (!totalExecutors.containsKey(componentName)) {
System.out.println("Our special-spout does not need scheduling.");
} else {
System.out.println("Our special-spout needs scheduling.");
List<ExecutorDetails> executors = totalExecutors.get(componentName);
// find out the our "special-supervisor" from the supervisor metadata
Collection<SupervisorDetails> supervisors = cluster.getSupervisors().values();
SupervisorDetails specialSupervisor = null;
for (SupervisorDetails supervisor : supervisors) {
Map meta = (Map) supervisor.getSchedulerMeta();
if(meta != null && meta.get("name") != null){
System.out.println("supervisor name:" + meta.get("name"));
if (meta.get("name").equals(supervisorName)) {
System.out.println("Supervisor finding");
specialSupervisor = supervisor;
break;
}
}else {
System.out.println("Supervisor meta null");
}
}
// found the special supervisor
if (specialSupervisor != null) {
System.out.println("Found the special-supervisor");
List<WorkerSlot> availableSlots = cluster.getAvailableSlots(specialSupervisor);
// 如果目标节点上已经没有空闲的slot,则进行强制释放
if (availableSlots.isEmpty() && !executors.isEmpty()) {
for (Integer port : cluster.getUsedPorts(specialSupervisor)) {
cluster.freeSlot(new WorkerSlot(specialSupervisor.getId(), port));
}
}
// 重新获取可用的slot
availableSlots = cluster.getAvailableSlots(specialSupervisor);
// 选取节点上第一个slot,进行分配
cluster.assign(availableSlots.get(0), topology.getId(), executors);
System.out.println("We assigned executors:" + executors + " to slot: [" + availableSlots.get(0).getNodeId() + ", " + availableSlots.get(0).getPort() + "]");
} else {
System.out.println("There is no supervisor find!!!");
}
}
}
}
说明部分
Storm自定义实现直接分配调度器,代码修改自Twitter Storm核心贡献者徐明明,此处为链接.
开发背景
在准备开发Storm自定义之前,事先已经了解了下现有Storm使用的调度器,默认是DefaultScheduler,调度原理大体如下:
* 在新的调度开始之前,先扫描一遍集群,如果有未释放掉的slot,则先进行释放
* 然后优先选择supervisor节点中有空闲的slot,进行分配,以达到最终平均分配资源的目标
现有scheduler的不足之处
上述的调度器基本可以满足一般要求,但是针对下面个例还是无法满足:
* 让spout分配到固定的机器上去,因为所需的数据就在那上面
* 不想让2个Topology运行在同一机器上,因为这2个Topology都很耗CPU
DirectScheduler的作用
DirectScheduler把划分单位缩小到组件级别,1个Spout和1个Bolt可以指定到某个节点上运行,如果没有指定,还是按照系统自带的调度器进行调度.这个配置在Topology提交的Conf配置中可配.
使用方法
集群配置
- 打包此项目,将jar包拷贝到STORM_HOME/lib目录下,在nimbus节点上的Storm包
在nimbus节点的storm.yaml配置中,进行如下的配置:
storm.scheduler: "storm.DirectScheduler"然后是在supervisor的节点中进行名称的配置,配置项如下:
supervisor.scheduler.meta:
name: "your-supervisor-name"
在集群这部分的配置就结束了,然后重启nimbus,supervisor节点即可,集群配置只要1次配置即可.
拓扑逻辑配置
见下面的代码设置,主要是把组件名和节点名称作为映射值传入
int numOfParallel;
TopologyBuilder builder;
StormTopology stormTopology;
Config config;
//待分配的组件名称与节点名称的映射关系
HashMap<String, String> component2Node;
//任务并行化数设为10个
numOfParallel = 2;
builder = new TopologyBuilder();
String desSpout = "my_spout";
String desBolt = "my_bolt";
//设置spout数据源
builder.setSpout(desSpout, new TestSpout(), numOfParallel);
builder.setBolt(desBolt, new TestBolt(), numOfParallel)
.shuffleGrouping(desSpout);
config = new Config();
config.setNumWorkers(numOfParallel);
config.setMaxSpoutPending(65536);
config.put(Config.STORM_ZOOKEEPER_CONNECTION_TIMEOUT, 40000);
config.put(Config.STORM_ZOOKEEPER_SESSION_TIMEOUT, 40000);
component2Node = new HashMap<>();
component2Node.put(desSpout, "special-supervisor1");
component2Node.put(desBolt, "special-supervisor2");
//此标识代表topology需要被调度
config.put("assigned_flag", "1");
//具体的组件节点对信息
config.put("design_map", component2Node);
StormSubmitter.submitTopology("test", config, builder.createTopology());
拓扑逻辑作业具体要被调度时,传入配置参数即可.
调度器后期优化
DirectScheduler只是针对原有的调度实现做了1层包装,后期可以进行更深层次的改造,涉及到节点在分配的时候slot的排序等等.
