YARN SLS-SLSRunner完整源码分析
整个slsruner从main函数开始：
public static void main(String[] argv) throws Exception {
exitAtTheFinish = true;
ToolRunner.run(new Configuration(), new SLSRunner(), argv);
函数中调用了run函数，并通过argv传递了命令行参数：
SLSrunner.java
public int run(final String[] argv) throws IOException, InterruptedException,
ParseException, ClassNotFoundException, YarnException {
Options options = new Options();
// Left for compatibility
options.addOption("inputrumen", true, "input rumen files");
options.addOption("inputsls", true, "input sls files");
// New more general format
options.addOption("tracetype", true, "the type of trace");
options.addOption("tracelocation", true, "input trace files");
options.addOption("nodes", true, "input topology");
options.addOption("output", true, "output directory");
options.addOption("trackjobs", true,
"jobs to be tracked during simulating");
options.addOption("printsimulation", false,
"print out simulation information");
CommandLineParser parser = new GnuParser();
CommandLine cmd = parser.parse(options, argv);
String traceType = null;
String traceLocation = null;
// compatibility with old commandline
if (cmd.hasOption("inputrumen")) {
traceType = "RUMEN";
traceLocation = cmd.getOptionValue("inputrumen");
}
if (cmd.hasOption("inputsls")) {
traceType = "SLS";
traceLocation = cmd.getOptionValue("inputsls");
}
if (cmd.hasOption("tracetype")) {
traceType = cmd.getOptionValue("tracetype");
traceLocation = cmd.getOptionValue("tracelocation");
}
String output = cmd.getOptionValue("output");
File outputFile = new File(output);
if (!outputFile.exists() && !outputFile.mkdirs()) {
System.err.println("ERROR: Cannot create output directory "
+ outputFile.getAbsolutePath());
throw new YarnException("Cannot create output directory");
}
Set<String> trackedJobSet = new HashSet<String>();
if (cmd.hasOption("trackjobs")) {
String trackjobs = cmd.getOptionValue("trackjobs");
String jobIds[] = trackjobs.split(",");
trackedJobSet.addAll(Arrays.asList(jobIds));
}
String tempNodeFile =
cmd.hasOption("nodes") ? cmd.getOptionValue("nodes") : "";
TraceType tempTraceType = TraceType.SLS;
switch (traceType) {
case "SLS":
tempTraceType = TraceType.SLS;
break;
case "RUMEN":
tempTraceType = TraceType.RUMEN;
break;
case "SYNTH":
tempTraceType = TraceType.SYNTH;
break;
default:
printUsage();
throw new YarnException("Misconfigured input");
}
String[] inputFiles = traceLocation.split(",");
setSimulationParams(tempTraceType, inputFiles, tempNodeFile, output,
trackedJobSet, cmd.hasOption("printsimulation"));
start();
return 0;
}首先是添加一些option，也就是保证能够识别这些参数，比如inputsls，nodes等。
添加好后，和传入的参数进行核对，得到被选中的参数及其值，比如：
// compatibility with old commandline
if (cmd.hasOption("inputrumen")) {
traceType = "RUMEN";
traceLocation = cmd.getOptionValue("inputrumen");
}
if (cmd.hasOption("inputsls")) {
traceType = "SLS";
traceLocation = cmd.getOptionValue("inputsls");
}这里如果传入的是SLS，那么traceType就是SLS，traceLocation就是inputsls参数传入的json文件了。tempTraceType就是TraceType.SLS。
函数最后调用start方法：
public void start() throws IOException, ClassNotFoundException, YarnException,
InterruptedException {
// start resource manager
startRM();
// start node managers
startNM();
// start application masters
startAM();
// set queue & tracked apps information
((SchedulerWrapper) rm.getResourceScheduler()).getTracker()
.setQueueSet(this.queueAppNumMap.keySet());
((SchedulerWrapper) rm.getResourceScheduler()).getTracker()
.setTrackedAppSet(this.trackedApps);
// print out simulation info
printSimulationInfo();
// blocked until all nodes RUNNING
waitForNodesRunning();
// starting the runner once everything is ready to go,
runner.start();
}start方法比较核心，在其中依次开启RM，NM，AM，然后等待所有Node正常运行后，开启runner。依次分析这些函数：
startRM：
private void startRM() throws ClassNotFoundException, YarnException {
Configuration rmConf = new YarnConfiguration(getConf());
String schedulerClass = rmConf.get(YarnConfiguration.RM_SCHEDULER);
if (Class.forName(schedulerClass) == CapacityScheduler.class) {
rmConf.set(YarnConfiguration.RM_SCHEDULER,
SLSCapacityScheduler.class.getName());
rmConf.setBoolean(YarnConfiguration.RM_SCHEDULER_ENABLE_MONITORS, true);
rmConf.set(YarnConfiguration.RM_SCHEDULER_MONITOR_POLICIES,
ProportionalCapacityPreemptionPolicy.class.getName());
} else if (Class.forName(schedulerClass) == FairScheduler.class) {
rmConf.set(YarnConfiguration.RM_SCHEDULER,
SLSFairScheduler.class.getName());
} else if (Class.forName(schedulerClass) == FifoScheduler.class) {
// TODO add support for FifoScheduler
throw new YarnException("Fifo Scheduler is not supported yet.");
}
rmConf.set(SLSConfiguration.METRICS_OUTPUT_DIR, metricsOutputDir);
final SLSRunner se = this;
rm = new ResourceManager() {
@Override
protected ApplicationMasterLauncher createAMLauncher() {
return new MockAMLauncher(se, this.rmContext, amMap);
}
};
// Across runs of parametrized tests, the JvmMetrics objects is retained,
// but is not registered correctly
JvmMetrics jvmMetrics = JvmMetrics.initSingleton("ResourceManager", null);
jvmMetrics.registerIfNeeded();
// Init and start the actual ResourceManager
rm.init(rmConf);
rm.start();
}首先是通过YarnConfiguration获得yarn-site.xml中的参数，然后从中获得schedule的方案，是capacity，FIFO或者fair。默认是fair。
然后创建一个ResourceManager的对象，并重写createAMLauncher方法，该方法是负责管理AM的，方法中调用了MockAMLauncher函数。该方法实现在MockAMLauncher.java中。
最后通过RM自身的rm.init和 rm.start开启RM。
startNM：
private void startNM() throws YarnException, IOException,
InterruptedException {
// nm configuration
int heartbeatInterval = getConf().getInt(
SLSConfiguration.NM_HEARTBEAT_INTERVAL_MS,
SLSConfiguration.NM_HEARTBEAT_INTERVAL_MS_DEFAULT);
float resourceUtilizationRatio = getConf().getFloat(
SLSConfiguration.NM_RESOURCE_UTILIZATION_RATIO,
SLSConfiguration.NM_RESOURCE_UTILIZATION_RATIO_DEFAULT);
// nm information (fetch from topology file, or from sls/rumen json file)
Map<String, Resource> nodeResourceMap = new HashMap<>();
Set<? extends String> nodeSet;
if (nodeFile.isEmpty()) {
for (String inputTrace : inputTraces) {
switch (inputType) {
case SLS:
nodeSet = SLSUtils.parseNodesFromSLSTrace(inputTrace);
for (String node : nodeSet) {
nodeResourceMap.put(node, null);
}
break;
case RUMEN:
nodeSet = SLSUtils.parseNodesFromRumenTrace(inputTrace);
for (String node : nodeSet) {
nodeResourceMap.put(node, null);
}
break;
case SYNTH:
stjp = new SynthTraceJobProducer(getConf(), new Path(inputTraces[0]));
nodeSet = SLSUtils.generateNodes(stjp.getNumNodes(),
stjp.getNumNodes()/stjp.getNodesPerRack());
for (String node : nodeSet) {
nodeResourceMap.put(node, null);
}
break;
default:
throw new YarnException("Input configuration not recognized, "
+ "trace type should be SLS, RUMEN, or SYNTH");
}
}
} else {
nodeResourceMap = SLSUtils.parseNodesFromNodeFile(nodeFile,
nodeManagerResource);
}
if (nodeResourceMap.size() == 0) {
throw new YarnException("No node! Please configure nodes.");
}
// create NM simulators
Random random = new Random();
Set<String> rackSet = new ConcurrentHashSet<>();
int threadPoolSize = Math.max(poolSize,
SLSConfiguration.RUNNER_POOL_SIZE_DEFAULT);
ExecutorService executorService = Executors.
newFixedThreadPool(threadPoolSize);
for (Map.Entry<String, Resource> entry : nodeResourceMap.entrySet()) {
executorService.submit(new Runnable() {
@Override public void run() {
try {
// we randomize the heartbeat start time from zero to 1 interval
NMSimulator nm = new NMSimulator();
Resource nmResource = nodeManagerResource;
String hostName = entry.getKey();
if (entry.getValue() != null) {
nmResource = entry.getValue();
}
nm.init(hostName, nmResource,
random.nextInt(heartbeatInterval),
heartbeatInterval, rm, resourceUtilizationRatio);
nmMap.put(nm.getNode().getNodeID(), nm);
runner.schedule(nm);
rackSet.add(nm.getNode().getRackName());
} catch (IOException | YarnException e) {
LOG.error("Got an error while adding node", e);
}
}
});
}
executorService.shutdown();
executorService.awaitTermination(10, TimeUnit.MINUTES);
numRacks = rackSet.size();
numNMs = nmMap.size();
}nodeResourceMap = SLSUtils.parseNodesFromNodeFile从--nodes参数获得的nodes机架配置文件nodeFile和nodeManagerResource作为参数，获得Node的所有参数。而nodeManagerResource是每个Node的mem，cpu信息。那么这些信息是什么时候从sls-runner.xml文件传入到nodeManagerResource的呢？
在SLSRunner.java中，有个init函数：
private void init(Configuration tempConf) throws ClassNotFoundException {
nmMap = new ConcurrentHashMap<>();
queueAppNumMap = new HashMap<>();
amMap = new ConcurrentHashMap<>();
amClassMap = new HashMap<>();
// runner configuration
setConf(tempConf);
// runner
poolSize = tempConf.getInt(SLSConfiguration.RUNNER_POOL_SIZE,
SLSConfiguration.RUNNER_POOL_SIZE_DEFAULT);
SLSRunner.runner.setQueueSize(poolSize);
// <AMType, Class> map
for (Map.Entry e : tempConf) {
String key = e.getKey().toString();
if (key.startsWith(SLSConfiguration.AM_TYPE_PREFIX)) {
String amType = key.substring(SLSConfiguration.AM_TYPE_PREFIX.length());
amClassMap.put(amType, Class.forName(tempConf.get(key)));
}
}
nodeManagerResource = getNodeManagerResource();
}该函数最后就从getNodeManagerResource初始化了nodeManagerResource。getNodeManagerResource方法如下：
private Resource getNodeManagerResource() {
Resource resource = Resources.createResource(0);
ResourceInformation[] infors = ResourceUtils.getResourceTypesArray();
for (ResourceInformation info : infors) {
long value;
if (info.getName().equals(ResourceInformation.MEMORY_URI)) {
value = getConf().getInt(SLSConfiguration.NM_MEMORY_MB,
SLSConfiguration.NM_MEMORY_MB_DEFAULT);
} else if (info.getName().equals(ResourceInformation.VCORES_URI)) {
value = getConf().getInt(SLSConfiguration.NM_VCORES,
SLSConfiguration.NM_VCORES_DEFAULT);
} else {
value = getConf().getLong(SLSConfiguration.NM_PREFIX +
info.getName(), SLSConfiguration.NM_RESOURCE_DEFAULT);
}
resource.setResourceValue(info.getName(), value);
}
return resource;
}可以看到这个方法就是从sls-runner.xml中获得每个Node的mem和cpu信息。
回到startNM函数：nodeResourceMap = SLSUtils.parseNodesFromNodeFile，获得信息后，开始创建NM simulator，调用nm.init来初始node的一些参数，比如hostName，nmResource（也就是nodeManagerResource：mem，cpu），然后加入到runner.schedule(nm); 也就是delayqueue中。
startAM：
private void startAM() throws YarnException, IOException {
switch (inputType) {
case SLS:
for (String inputTrace : inputTraces) {
startAMFromSLSTrace(inputTrace);
}
break;
case RUMEN:
long baselineTimeMS = 0;
for (String inputTrace : inputTraces) {
startAMFromRumenTrace(inputTrace, baselineTimeMS);
}
break;
case SYNTH:
startAMFromSynthGenerator();
break;
default:
throw new YarnException("Input configuration not recognized, "
+ "trace type should be SLS, RUMEN, or SYNTH");
}
numAMs = amMap.size();
remainingApps = numAMs;
}这里就是一个选择输入的格式会有对应的函数，这里看下sls：startAMFromSLSTrace
/**
* Parse workload from a SLS trace file.
*/
@SuppressWarnings("unchecked")
private void startAMFromSLSTrace(String inputTrace) throws IOException {
JsonFactory jsonF = new JsonFactory();
ObjectMapper mapper = new ObjectMapper();
try (Reader input = new InputStreamReader(
new FileInputStream(inputTrace), "UTF-8")) {
Iterator<Map> jobIter = mapper.readValues(
jsonF.createParser(input), Map.class);
while (jobIter.hasNext()) {
try {
createAMForJob(jobIter.next());
} catch (Exception e) {
LOG.error("Failed to create an AM: {}", e.getMessage());
}
}
}
}从--input-sls参数后的文件中获得workload信息，也就是jobh和task信息。
while (jobIter.hasNext()) {
try {
createAMForJob(jobIter.next());
} catch (Exception e) {
LOG.error("Failed to create an AM: {}", e.getMessage());
}
}while是循环看有多少个job，若通过官方给出的例子，应该是有两个job。每一个job会调用createAMForJob函数：
private void createAMForJob(Map jsonJob) throws YarnException {
long jobStartTime = Long.parseLong(
jsonJob.get(SLSConfiguration.JOB_START_MS).toString());
long jobFinishTime = 0;
if (jsonJob.containsKey(SLSConfiguration.JOB_END_MS)) {
jobFinishTime = Long.parseLong(
jsonJob.get(SLSConfiguration.JOB_END_MS).toString());
}
String user = (String) jsonJob.get(SLSConfiguration.JOB_USER);
if (user == null) {
user = "default";
}
String queue = jsonJob.get(SLSConfiguration.JOB_QUEUE_NAME).toString();
increaseQueueAppNum(queue);
String amType = (String)jsonJob.get(SLSConfiguration.AM_TYPE);
if (amType == null) {
amType = SLSUtils.DEFAULT_JOB_TYPE;
}
int jobCount = 1;
if (jsonJob.containsKey(SLSConfiguration.JOB_COUNT)) {
jobCount = Integer.parseInt(
jsonJob.get(SLSConfiguration.JOB_COUNT).toString());
}
jobCount = Math.max(jobCount, 1);
String oldAppId = (String)jsonJob.get(SLSConfiguration.JOB_ID);
// Job id is generated automatically if this job configuration allows
// multiple job instances
if(jobCount > 1) {
oldAppId = null;
}
for (int i = 0; i < jobCount; i++) {
runNewAM(amType, user, queue, oldAppId, jobStartTime, jobFinishTime,
getTaskContainers(jsonJob), getAMContainerResource(jsonJob));
}
}从sls-jobs.json中获得jobStartTime，jobFinishTime。user，queue，amType等信息。
然后调用runNewAM函数开启AM，将上面的参数信息传递过去。runNewAM还有其他两个参数：getTaskContainers(jsonJob), getAMContainerResource(jsonJob)。
getTaskContainers：
private List<ContainerSimulator> getTaskContainers(Map jsonJob)
throws YarnException {
List<ContainerSimulator> containers = new ArrayList<>();
List tasks = (List) jsonJob.get(SLSConfiguration.JOB_TASKS);
if (tasks == null || tasks.size() == 0) {
throw new YarnException("No task for the job!");
}
for (Object o : tasks) {
Map jsonTask = (Map) o;
String hostname = (String) jsonTask.get(SLSConfiguration.TASK_HOST);
long duration = 0;
if (jsonTask.containsKey(SLSConfiguration.TASK_DURATION_MS)) {
duration = Integer.parseInt(
jsonTask.get(SLSConfiguration.TASK_DURATION_MS).toString());
} else if (jsonTask.containsKey(SLSConfiguration.DURATION_MS)) {
// Also support "duration.ms" for backward compatibility
duration = Integer.parseInt(
jsonTask.get(SLSConfiguration.DURATION_MS).toString());
} else if (jsonTask.containsKey(SLSConfiguration.TASK_START_MS) &&
jsonTask.containsKey(SLSConfiguration.TASK_END_MS)) {
long taskStart = Long.parseLong(
jsonTask.get(SLSConfiguration.TASK_START_MS).toString());
long taskFinish = Long.parseLong(
jsonTask.get(SLSConfiguration.TASK_END_MS).toString());
duration = taskFinish - taskStart;
}
if (duration <= 0) {
throw new YarnException("Duration of a task shouldn't be less or equal"
+ " to 0!");
}
Resource res = getResourceForContainer(jsonTask);
int priority = DEFAULT_MAPPER_PRIORITY;
if (jsonTask.containsKey(SLSConfiguration.TASK_PRIORITY)) {
priority = Integer.parseInt(
jsonTask.get(SLSConfiguration.TASK_PRIORITY).toString());
}
String type = "map";
if (jsonTask.containsKey(SLSConfiguration.TASK_TYPE)) {
type = jsonTask.get(SLSConfiguration.TASK_TYPE).toString();
}
int count = 1;
if (jsonTask.containsKey(SLSConfiguration.COUNT)) {
count = Integer.parseInt(
jsonTask.get(SLSConfiguration.COUNT).toString());
}
count = Math.max(count, 1);
ExecutionType executionType = ExecutionType.GUARANTEED;
if (jsonTask.containsKey(SLSConfiguration.TASK_EXECUTION_TYPE)) {
executionType = ExecutionType.valueOf(
jsonTask.get(SLSConfiguration.TASK_EXECUTION_TYPE).toString());
}
for (int i = 0; i < count; i++) {
containers.add(
new ContainerSimulator(res, duration, hostname, priority, type,
executionType));
}
}
return containers;
}该函数是将sls-jobs.json中某个具体jobs下的task任务以container的形式分配出来。
List<ContainerSimulator> containers = new ArrayList<>();函数返回的值就是List类型的containers，也就是有多少个task，list就有多大。
List tasks = (List) jsonJob.get(SLSConfiguration.JOB_TASKS);
if (tasks == null || tasks.size() == 0) {
throw new YarnException("No task for the job!");
}
for (Object o : tasks) {
Map jsonTask = (Map) o;
String hostname = (String) jsonTask.get(SLSConfiguration.TASK_HOST);
long duration = 0;
if (jsonTask.containsKey(SLSConfiguration.TASK_DURATION_MS)) {
duration = Integer.parseInt(
jsonTask.get(SLSConfiguration.TASK_DURATION_MS).toString());
} else if (jsonTask.containsKey(SLSConfiguration.DURATION_MS)) {
// Also support "duration.ms" for backward compatibility
duration = Integer.parseInt(
jsonTask.get(SLSConfiguration.DURATION_MS).toString());
} else if (jsonTask.containsKey(SLSConfiguration.TASK_START_MS) &&
jsonTask.containsKey(SLSConfiguration.TASK_END_MS)) {
long taskStart = Long.parseLong(
jsonTask.get(SLSConfiguration.TASK_START_MS).toString());
long taskFinish = Long.parseLong(
jsonTask.get(SLSConfiguration.TASK_END_MS).toString());
duration = taskFinish - taskStart;
}
if (duration <= 0) {
throw new YarnException("Duration of a task shouldn't be less or equal"
+ " to 0!");
}
Resource res = getResourceForContainer(jsonTask);
int priority = DEFAULT_MAPPER_PRIORITY;
if (jsonTask.containsKey(SLSConfiguration.TASK_PRIORITY)) {
priority = Integer.parseInt(
jsonTask.get(SLSConfiguration.TASK_PRIORITY).toString());
}
String type = "map";
if (jsonTask.containsKey(SLSConfiguration.TASK_TYPE)) {
type = jsonTask.get(SLSConfiguration.TASK_TYPE).toString();
}
int count = 1;
if (jsonTask.containsKey(SLSConfiguration.COUNT)) {
count = Integer.parseInt(
jsonTask.get(SLSConfiguration.COUNT).toString());
}
count = Math.max(count, 1);
ExecutionType executionType = ExecutionType.GUARANTEED;
if (jsonTask.containsKey(SLSConfiguration.TASK_EXECUTION_TYPE)) {
executionType = ExecutionType.valueOf(
jsonTask.get(SLSConfiguration.TASK_EXECUTION_TYPE).toString());
}
for (int i = 0; i < count; i++) {
containers.add(
new ContainerSimulator(res, duration, hostname, priority, type,
executionType));
}
}然后将tasks一一取出，通过for循环来为每一个分配container。
在循环中，先是获得task的信息，比如start，end 时间，优先级，task类型（map）。然后将这些信息收集起来，传递给getResourceForContainer函数，来申请container：
Resource res = getResourceForContainer(jsonTask); private Resource getResourceForContainer(Map jsonTask) {
Resource res = getDefaultContainerResource();
ResourceInformation[] infors = ResourceUtils.getResourceTypesArray();
for (ResourceInformation info : infors) {
if (jsonTask.containsKey(SLSConfiguration.TASK_PREFIX + info.getName())) {
long value = Long.parseLong(
jsonTask.get(SLSConfiguration.TASK_PREFIX + info.getName())
.toString());
res.setResourceValue(info.getName(), value);
}
}
return res;
}循环最后，将申请的资源res加入到待分配container list中：
for (int i = 0; i < count; i++) {
containers.add(
new ContainerSimulator(res, duration, hostname, priority, type,
executionType));
}跳出循环，最后返回containers。也就是作为runNewAM的参数了。
接下来是getAMContainerResource-runNewAM另一个参数：
private Resource getAMContainerResource(Map jsonJob) {
Resource amContainerResource =
SLSConfiguration.getAMContainerResource(getConf());
if (jsonJob == null) {
return amContainerResource;
}
ResourceInformation[] infors = ResourceUtils.getResourceTypesArray();
for (ResourceInformation info : infors) {
String key = SLSConfiguration.JOB_AM_PREFIX + info.getName();
if (jsonJob.containsKey(key)) {
long value = Long.parseLong(jsonJob.get(key).toString());
amContainerResource.setResourceValue(info.getName(), value);
}
}
return amContainerResource;
}这段就是为AM分配资源也就是container。
然后回到runNewAM函数，也就是为每一个jobs分配了一个AM的资源以及jobs下所有tasks的资源。
runNewAM：
private void runNewAM(String jobType, String user,
String jobQueue, String oldJobId, long jobStartTimeMS,
long jobFinishTimeMS, List<ContainerSimulator> containerList,
ReservationId reservationId, long deadline, Resource amContainerResource,
Map<String, String> params) {
AMSimulator amSim = (AMSimulator) ReflectionUtils.newInstance(
amClassMap.get(jobType), new Configuration());
if (amSim != null) {
int heartbeatInterval = getConf().getInt(
SLSConfiguration.AM_HEARTBEAT_INTERVAL_MS,
SLSConfiguration.AM_HEARTBEAT_INTERVAL_MS_DEFAULT);
boolean isTracked = trackedApps.contains(oldJobId);
if (oldJobId == null) {
oldJobId = Integer.toString(AM_ID);
}
AM_ID++;
amSim.init(heartbeatInterval, containerList, rm, this, jobStartTimeMS,
jobFinishTimeMS, user, jobQueue, isTracked, oldJobId,
runner.getStartTimeMS(), amContainerResource, params);
if(reservationId != null) {
// if we have a ReservationId, delegate reservation creation to
// AMSim (reservation shape is impl specific)
UTCClock clock = new UTCClock();
amSim.initReservation(reservationId, deadline, clock.getTime());
}
runner.schedule(amSim);
maxRuntime = Math.max(maxRuntime, jobFinishTimeMS);
numTasks += containerList.size();
amMap.put(oldJobId, amSim);
}
}、创建了AMSimulator对象。amSim.init初始化该AMSIM，然后通过runner.schedule加入到delayqueue中，和NM一样。
初始化完RM，NM，AM后，调用runner.start();开始运行：
TaskRunner.java:
@SuppressWarnings("unchecked")
public void start() {
if (executor != null && !executor.isTerminated()) {
throw new IllegalStateException("Executor already running");
}
DelayQueue preStartQueue = queue;
queue = new DelayQueue();
executor = new ThreadPoolExecutor(threadPoolSize, threadPoolSize, 0,
TimeUnit.MILLISECONDS, queue);
executor.prestartAllCoreThreads();
startTimeMS = System.currentTimeMillis();
for (Object d : preStartQueue) {
schedule((Task) d, startTimeMS);
}
}也就是都放入schedule进行调度，循环运行了。
在java中，掉用完一个类的start函数后，会自动调用该类的run方法：
@Override
public final void run() {
try {
if (nextRun == startTime) {
firstStep();
nextRun += repeatInterval;
if (nextRun <= endTime) {
queue.add(this);
}
} else if (nextRun < endTime) {
middleStep();
nextRun += repeatInterval;
queue.add(this);
} else {
lastStep();
}
} catch (Exception e) {
e.printStackTrace();
Thread.getDefaultUncaughtExceptionHandler()
.uncaughtException(Thread.currentThread(), e);
}
}也就是循环运行run，在run中，检查nextRun的值，来确定task执行到哪一步了，并执行对应的函数。
其中firststep，middlestep，laststep都被AM，NM这些类重载了，也就是会调用到对应的函数。
比如如果是AM的task执行，首先执行firststep：
AMSimulator.java
/**
* register with RM
*/
@Override
public void firstStep() throws Exception {
simulateStartTimeMS = System.currentTimeMillis() - baselineTimeMS;
ReservationId reservationId = null;
// submit a reservation if one is required, exceptions naturally happen
// when the reservation does not fit, catch, log, and move on running job
// without reservation.
try {
reservationId = submitReservationWhenSpecified();
} catch (UndeclaredThrowableException y) {
LOG.warn("Unable to place reservation: " + y.getMessage());
}
// submit application, waiting until ACCEPTED
submitApp(reservationId);
// track app metrics
trackApp();
}这一步是向RM注册，并提交APP（submitApp），其实这有点类似于YARN中client的行为。在SLS中都简化给AM做了。
也就是模拟client端向RM发送app信息。
再来看middlestep：
@Override
public void middleStep() throws Exception {
if (isAMContainerRunning) {
// process responses in the queue
processResponseQueue();
// send out request
sendContainerRequest();
// check whether finish
checkStop();
}
}首先判断AM有没有运行起来，如果运行起来就继续执行：
processResponseQueue函数是检测queue中container的运行状态，包括完成的，未完成的，刚分配的等。
sendContainerRequest是发送请求，也就是发送job中task的具体请求，是map还是reduce
sendContainerRequest：MRAMSimulator.java
@Override
protected void sendContainerRequest()
throws YarnException, IOException, InterruptedException {
if (isFinished) {
return;
}
// send out request
List<ResourceRequest> ask = null;
if (mapFinished != mapTotal) {
// map phase
if (!pendingMaps.isEmpty()) {
ask = packageRequests(mergeLists(pendingMaps, scheduledMaps),
PRIORITY_MAP);
LOG.debug("Application {} sends out request for {} mappers.",
appId, pendingMaps.size());
scheduledMaps.addAll(pendingMaps);
pendingMaps.clear();
} else if (!pendingFailedMaps.isEmpty()) {
ask = packageRequests(mergeLists(pendingFailedMaps, scheduledMaps),
PRIORITY_MAP);
LOG.debug("Application {} sends out requests for {} failed mappers.",
appId, pendingFailedMaps.size());
scheduledMaps.addAll(pendingFailedMaps);
pendingFailedMaps.clear();
}
} else if (reduceFinished != reduceTotal) {
// reduce phase
if (!pendingReduces.isEmpty()) {
ask = packageRequests(mergeLists(pendingReduces, scheduledReduces),
PRIORITY_REDUCE);
LOG.debug("Application {} sends out requests for {} reducers.",
appId, pendingReduces.size());
scheduledReduces.addAll(pendingReduces);
pendingReduces.clear();
} else if (!pendingFailedReduces.isEmpty()) {
ask = packageRequests(mergeLists(pendingFailedReduces, scheduledReduces),
PRIORITY_REDUCE);
LOG.debug("Application {} sends out request for {} failed reducers.",
appId, pendingFailedReduces.size());
scheduledReduces.addAll(pendingFailedReduces);
pendingFailedReduces.clear();
}
}
if (ask == null) {
ask = new ArrayList<>();
}
final AllocateRequest request = createAllocateRequest(ask);
if (totalContainers == 0) {
request.setProgress(1.0f);
} else {
request.setProgress((float) finishedContainers / totalContainers);
}
UserGroupInformation ugi =
UserGroupInformation.createRemoteUser(appAttemptId.toString());
Token<AMRMTokenIdentifier> token = rm.getRMContext().getRMApps()
.get(appAttemptId.getApplicationId())
.getRMAppAttempt(appAttemptId).getAMRMToken();
ugi.addTokenIdentifier(token.decodeIdentifier());
AllocateResponse response = ugi.doAs(
new PrivilegedExceptionAction<AllocateResponse>() {
@Override
public AllocateResponse run() throws Exception {
return rm.getApplicationMasterService().allocate(request);
}
});
if (response != null) {
responseQueue.put(response);
}
}将task对容器的请求通过packageRequests函数打包成ask。
packageRequests函数是对数据包进行打包的最重要的方法，也就是协议的打包。涉及到协议的传输。会专门写篇文章详细描述。
然后通过request = createAllocateRequest(ask);生成标准格式的request（AllocateRequest类），然后发送给RM：
UserGroupInformation ugi =
UserGroupInformation.createRemoteUser(appAttemptId.toString());
Token<AMRMTokenIdentifier> token = rm.getRMContext().getRMApps()
.get(appAttemptId.getApplicationId())
.getRMAppAttempt(appAttemptId).getAMRMToken();
ugi.addTokenIdentifier(token.decodeIdentifier());
AllocateResponse response = ugi.doAs(
new PrivilegedExceptionAction<AllocateResponse>() {
@Override
public AllocateResponse run() throws Exception {
return rm.getApplicationMasterService().allocate(request);
}
});
if (response != null) {
responseQueue.put(response);得到的回复是response（AllocateResponse类），然后放入responseQueue中。完成申请过程。
可以看到请求：request（AllocateRequest类），回复：response（AllocateResponse类）
都是在org.apache.hadoop.yarn.api.protocolrecords.下的类也就是协议。
laststep：
@Override
public void lastStep() throws Exception {
LOG.info("Application {} is shutting down.", appId);
// unregister tracking
if (isTracked) {
untrackApp();
}
// Finish AM container
if (amContainer != null) {
LOG.info("AM container = {} reported to finish", amContainer.getId());
se.getNmMap().get(amContainer.getNodeId()).cleanupContainer(
amContainer.getId());
} else {
LOG.info("AM container is null");
}
if (null == appAttemptId) {
// If appAttemptId == null, AM is not launched from RM's perspective, so
// it's unnecessary to finish am as well
return;
}
// unregister application master
final FinishApplicationMasterRequest finishAMRequest = recordFactory
.newRecordInstance(FinishApplicationMasterRequest.class);
finishAMRequest.setFinalApplicationStatus(FinalApplicationStatus.SUCCEEDED);
UserGroupInformation ugi =
UserGroupInformation.createRemoteUser(appAttemptId.toString());
Token<AMRMTokenIdentifier> token = rm.getRMContext().getRMApps().get(appId)
.getRMAppAttempt(appAttemptId).getAMRMToken();
ugi.addTokenIdentifier(token.decodeIdentifier());
ugi.doAs(new PrivilegedExceptionAction<Object>() {
@Override
public Object run() throws Exception {
rm.getApplicationMasterService()
.finishApplicationMaster(finishAMRequest);
return null;
}
});
simulateFinishTimeMS = System.currentTimeMillis() - baselineTimeMS;
// record job running information
SchedulerMetrics schedulerMetrics =
((SchedulerWrapper)rm.getResourceScheduler()).getSchedulerMetrics();
if (schedulerMetrics != null) {
schedulerMetrics.addAMRuntime(appId, traceStartTimeMS, traceFinishTimeMS,
simulateStartTimeMS, simulateFinishTimeMS);
}
}首先结束掉AM的container，然后远程将RM的AM停掉：finishApplicationMaster(finishAMRequest)。
这就完成了整个slsrunner。回顾源码可以发现：从runner.start()开始，其实程序就不在slsrunner.java中了，一直在schedule中调度循环。
posted on 2018-07-15 11:10 sichenzhao 阅读(383) 评论(0) 编辑 收藏 举报
