一、任务调度策略
当我们将topology提交到storm集群的时候,任务是怎样分配的呢,这就需要理解storm的任务调度策略,这里主要给大家分享默认的调度策略DefaultScheduler,在storm的1.1.0版本已经支持4种调度策略,分别是DefaultScheduler,IsolationScheduler,MultitenantScheduler,ResourceAwareScheduler。
二、Topology的提交过程
在理解默认的调度策略之前,先看一下我们提交一个topology到集群的整个流程图。
主要分为几步:
1、非本地模式下,客户端通过thrift调用nimbus接口,来上传代码到nimbus并触发提交操作.
2、nimbus进行任务分配,并将信息同步到zookeeper.
3、supervisor定期获取任务分配信息,如果topology代码缺失,会从nimbus下载代码,并根据任务分配信息,同步worker.
4、worker根据分配的tasks信息,启动多个executor线程,同时实例化spout、bolt、acker等组件,此时,等待所有connections(worker和其它机器通讯的网络连接)启动完毕,storm集群即进入工作状态。
5、除非显示调用kill topology,否则spout、bolt等组件会一直运行。
下面我们来看一下整个topolgoy提交过程的源代码
Main方法里面的提交代码
StormSubmitter.submitTopology("one-work",config,builder.createTopology());
然后调用下面方法
public static void submitTopologyAs(String name, Map stormConf, StormTopology topology, SubmitOptions opts, ProgressListener progressListener, String asUser)
throws AlreadyAliveException, InvalidTopologyException, AuthorizationException, IllegalArgumentException {
//配置文件必须能够被Json序列化
if(!Utils.isValidConf(stormConf)) {
throw new IllegalArgumentException("Storm conf is not valid. Must be json-serializable");
}
stormConf = new HashMap(stormConf);
//将命令行的参数加入stormConf
stormConf.putAll(Utils.readCommandLineOpts());
//先加载defaults.yaml, 然后再加载storm.yaml
Map conf = Utils.readStormConfig();
conf.putAll(stormConf);
//设置zookeeper的相关权限
stormConf.putAll(prepareZookeeperAuthentication(conf));
validateConfs(conf, topology);
Map<String,String> passedCreds = new HashMap<>();
if (opts != null) {
Credentials tmpCreds = opts.get_creds();
if (tmpCreds != null) {
passedCreds = tmpCreds.get_creds();
}
}
Map<String,String> fullCreds = populateCredentials(conf, passedCreds);
if (!fullCreds.isEmpty()) {
if (opts == null) {
opts = new SubmitOptions(TopologyInitialStatus.ACTIVE);
}
opts.set_creds(new Credentials(fullCreds));
}
try {
//本地模式
if(localNimbus!=null) {
LOG.info("Submitting topology " + name + " in local mode");
if(opts!=null) {
localNimbus.submitTopologyWithOpts(name, stormConf, topology, opts);
} else {
// this is for backwards compatibility
localNimbus.submitTopology(name, stormConf, topology);
}
LOG.info("Finished submitting topology: " + name);
//这里重点分析将topology提交到集群模式
} else {
//将配置信息转为json字符串
String serConf = JSONValue.toJSONString(stormConf);
//校验集群中topology-name是否已经存在
if(topologyNameExists(conf, name, asUser)) {
throw new RuntimeException("Topology with name `" + name + "` already exists on cluster");
}
//将jar包上传至nimbus,这个时候topology还没有正在跑起来,只是将jar提交到了nimbus,等待后续的任务调度
String jar = submitJarAs(conf, System.getProperty("storm.jar"), progressListener, asUser);
try (
//获取Nimbus client对象
NimbusClient client = NimbusClient.getConfiguredClientAs(conf, asUser)){
LOG.info("Submitting topology " + name + " in distributed mode with conf " + serConf);
//调用submitTopologyWithOpts正式向nimbus提交拓扑,其实所谓的提交拓扑,就是将拓扑的配置信息通过thrift发送到thrift server,并把jar包上传到nimbus,等待nimbus的后续处//理,此时拓扑并未真正起来,直至recv_submitTopology获得成功的返回信息为止
if (opts != null) {
client.getClient().submitTopologyWithOpts(name, jar, serConf, topology, opts);
} else {
// this is for backwards compatibility
client.getClient().submitTopology(name, jar, serConf, topology);
}
LOG.info("Finished submitting topology: " + name);
} catch (InvalidTopologyException e) {
LOG.warn("Topology submission exception: " + e.get_msg());
throw e;
} catch (AlreadyAliveException e) {
LOG.warn("Topology already alive exception", e);
throw e;
}
}
} catch(TException e) {
throw new RuntimeException(e);
}
invokeSubmitterHook(name, asUser, conf, topology);
}
继续调用
public static String submitJarAs(Map conf, String localJar, ProgressListener listener, String asUser) {
if (localJar == null) {
throw new RuntimeException("Must submit topologies using the 'storm' client script so that StormSubmitter knows which jar to upload.");
}
//如果获取了nimbus client
try (NimbusClient client = NimbusClient.getConfiguredClientAs(conf, asUser)) {
//获取topology-jar对应的存放地址
String uploadLocation = client.getClient().beginFileUpload();
LOG.info("Uploading topology jar " + localJar + " to assigned location: " + uploadLocation);
BufferFileInputStream is = new BufferFileInputStream(localJar, THRIFT_CHUNK_SIZE_BYTES);
long totalSize = new File(localJar).length();
if (listener != null) {
listener.onStart(localJar, uploadLocation, totalSize);
}
long bytesUploaded = 0;
while(true) {
byte[] toSubmit = is.read();
bytesUploaded += toSubmit.length;
if (listener != null) {
listener.onProgress(localJar, uploadLocation, bytesUploaded, totalSize);
}
if(toSubmit.length==0) break;
//一块一块的提交jar
client.getClient().uploadChunk(uploadLocation, ByteBuffer.wrap(toSubmit));
}
//完成jar包提交
client.getClient().finishFileUpload(uploadLocation);
if (listener != null) {
listener.onCompleted(localJar, uploadLocation, totalSize);
}
LOG.info("Successfully uploaded topology jar to assigned location: " + uploadLocation);
//返回存放jar的位置
return uploadLocation;
} catch(Exception e) {
throw new RuntimeException(e);
}
}
继续调用
public void submitTopology(String name, String uploadedJarLocation, String jsonConf, StormTopology topology) throws AlreadyAliveException, InvalidTopologyException, AuthorizationException, org.apache.thrift.TException
{
//发送topology相关信息到nimbus
send_submitTopology(name, uploadedJarLocation, jsonConf, topology);
//接收返回结果
recv_submitTopology();
}
继续调用:
public void send_submitTopology(String name, String uploadedJarLocation, String jsonConf, StormTopology topology) throws org.apache.thrift.TException{
submitTopology_args args = new submitTopology_args();
args.set_name(name);
args.set_uploadedJarLocation(uploadedJarLocation);
args.set_jsonConf(jsonConf);
args.set_topology(topology);
sendBase("submitTopology", args);
}
继续调用:
public void recv_submitTopology() throws AlreadyAliveException, InvalidTopologyException, AuthorizationException, org.apache.thrift.TException
{
submitTopology_result result = new submitTopology_result();
receiveBase(result, "submitTopology");
if (result.e != null) {
throw result.e;
}
if (result.ite != null) {
throw result.ite;
}
if (result.aze != null) {
throw result.aze;
}
return;
}
三、任务分配
在上面我们已经将topology提交到到nimbus了,下一步就是任务分配,strom默认4种分配策略。
DefaultScheduler策略,DefaultScheduler其实主要有几步
1、首先是获取当前集群中需要进行任务分配的topology
2、获取整个集群可用的slot
3、获取当前topology需要分配的executor信息
4、计算当前集群可释放的slot
5、统计可释放的solt和空闲的solt
6、执行topology分配
下面我们用一个列子来说明
比如初始状态下,集群的状态如下:2个supervisor,每个supervisor有4个可用的端口,这里我已A,B分别代表2个supervisor,那么初始情况下整个集群可用的端口地址就是:
A-6700,A-6701,A-6703,A-6704,B-6700,B-6701,B6702,B-6703。
Step1:现在我提交一个topology到集群,这个拓扑我给他分配2个worker端口,6个executor线程,每个线程默认运行一个任务就是6个task。当我们提交这个拓扑的时候,首先集群会将可用的solts进行排序如上可用端口的顺序,然后计算线程和任务的对应关系,这里都是6个,格式为[start-task-id end-task-id]就[1,1][2,2][3,3],[4,4],[5,5],[6,6]然后分配到2个worker上,那么每个worker分别跑3个线程即分配状态为[3,3]。
综上:分配的结果为:
[1,1],[2,2],[3,3] --->worker1
[4,4],[5,5],[6,6] --->worker2
而非常重要的是storm为了合理利用资源,在将可用slots排序后,依次选择worker来运行任务,也就是worker1对应A--6700,worker2对应B--6700。
下面我们来看一下storm集群的日志文件
首先提交topology
然后看一下nimbus.log日志
2017-04-09 22:00:12.502 o.a.s.d.common [INFO] Started statistics report plugin...
2017-04-09 22:00:12.575 o.a.s.d.nimbus [INFO] Starting nimbus server for storm version '1.0.0'
2017-04-09 22:03:13.661 o.a.s.d.nimbus [INFO] Uploading file from client to /bigdata/storm/datas/nimbus/inbox/stormjar-f16a2908-869a-418d-a589-ff6c7968724f.jar
2017-04-09 22:03:16.163 o.a.s.d.nimbus [INFO] Finished uploading file from client: /bigdata/storm/datas/nimbus/inbox/stormjar-f16a2908-869a-418d-a589-ff6c7968724f.jar
2017-04-09 22:03:16.328 o.a.s.d.nimbus [INFO] Received topology submission for testTopologySubmit with conf {"topology.max.task.parallelism" nil, "topology.submitter.principal" "", "topology.acker.executors" nil, "topology.eventlogger.executors" 0, "topology.workers" 2, "topology.debug" false, "storm.zookeeper.superACL" nil, "topology.users" (), "topology.submitter.user" "root", "topology.kryo.register" nil, "topology.kryo.decorators" (), "storm.id" "testTopologySubmit-1-1491800596", "topology.name" "testTopologySubmit"}
2017-04-09 22:03:16.335 o.a.s.d.nimbus [INFO] uploadedJar /bigdata/storm/datas/nimbus/inbox/stormjar-f16a2908-869a-418d-a589-ff6c7968724f.jar
获取集群可用的solts:
可以看到分配到了slave1和slave2的6700端口
slave1--132机器
slave2-134机器
Step2:现在整个集群还有A-6701,A-6702,A-6703,B-6701,B-6702,B-6703,现在假如我要提交一个新的topology,然后只有1个worker,那么它会分配到A-6701,那么如果后面每次都提交只需要一个worker的topology,那么会导致A机器端口已经被分配完了,而B机器还有3个可用的端口,所有storm的任务调度也不是很公平的,A机器已经满载了,B机器还有3个可用端口。