[
https://issues.apache.org/jira/browse/FLINK-8707?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16406934#comment-16406934
]
Stephan Ewen commented on FLINK-8707:
-------------------------------------
Thanks for all the additional data.
My first thoughts are that this could have indeed something to do with the
classloading, or potentially some auxiliary service (rest / metrics, etc).
Brief recap on what Flink does classloading wise:
- JobManager and TaskManager start with the classes in {{lib}} in the
classpath.
- Whenever a task is started in the TaskManager, the TaskManager ensures it
has a local cached copy of all JAR files belonging to that job, and creates a
dedicated classloader for that task. When the task is removed (finished,
canceled, failed), the classloader is closed. For the URL Classloader that
Flink uses, that is supposed to release the file references.
So, indeed, the jars could be referenced by multiple classloaders. But the
factor you see is excessive, I cannot really imagine why so many references to
the files should appear.
To debug this further, could you check the following:
- When you start a set of Flink processes, do the TaskManagers immediately
have many FDs?
- When you start the job, do the TaskManagers immediately go high with the
FDs?
- When the job fails/recovers, does each recovery add additional open FDs?
- When the job does not fail but just keeps running, do FDs grow steadily
over time?
Thanks for helping with this one!
> Excessive amount of files opened by flink task manager
> ------------------------------------------------------
>
> Key: FLINK-8707
> URL: https://issues.apache.org/jira/browse/FLINK-8707
> Project: Flink
> Issue Type: Bug
> Components: TaskManager
> Affects Versions: 1.3.2
> Environment: NAME="Red Hat Enterprise Linux Server"
> VERSION="7.3 (Maipo)"
> Two boxes, each with a Job Manager & Task Manager, using Zookeeper for HA.
> flink.yaml below with some settings (removed exact box names) etc:
> env.log.dir: ...some dir...residing on the same box
> env.pid.dir: some dir...residing on the same box
> metrics.reporter.jmx.class: org.apache.flink.metrics.jmx.JMXReporter
> metrics.reporters: jmx
> state.backend: filesystem
> state.backend.fs.checkpointdir: file:///some_nfs_mount
> state.checkpoints.dir: file:///some_nfs_mount
> state.checkpoints.num-retained: 3
> high-availability.cluster-id: /tst
> high-availability.storageDir: file:///some_nfs_mount/ha
> high-availability: zookeeper
> high-availability.zookeeper.path.root: /flink
> high-availability.zookeeper.quorum: ...list of zookeeper boxes
> env.java.opts.jobmanager: ...some extra jar args
> jobmanager.archive.fs.dir: some dir...residing on the same box
> jobmanager.web.submit.enable: true
> jobmanager.web.tmpdir: some dir...residing on the same box
> env.java.opts.taskmanager: some extra jar args
> taskmanager.tmp.dirs: some dir...residing on the same box/var/tmp
> taskmanager.network.memory.min: 1024MB
> taskmanager.network.memory.max: 2048MB
> blob.storage.directory: some dir...residing on the same box
> Reporter: Alexander Gardner
> Priority: Blocker
> Fix For: 1.5.0
>
> Attachments: AfterRunning-3-jobs-Box2-TM-JCONSOLE.png,
> AfterRunning-3-jobs-TM-FDs-BOX2.jpg, AfterRunning-3-jobs-lsof-p.box2-TM,
> AfterRunning-3-jobs-lsof.box2-TM, AterRunning-3-jobs-Box1-TM-JCONSOLE.png,
> box1-jobmgr-lsof, box1-taskmgr-lsof, box2-jobmgr-lsof, box2-taskmgr-lsof
>
>
> The job manager has less FDs than the task manager.
>
> Hi
> A support alert indicated that there were a lot of open files for the boxes
> running Flink.
> There were 4 flink jobs that were dormant but had consumed a number of msgs
> from Kafka using the FlinkKafkaConsumer010.
> A simple general lsof:
> $ lsof | wc -l -> returned 153114 open file descriptors.
> Focusing on the TaskManager process (process ID = 12154):
> $ lsof | grep 12154 | wc -l- > returned 129322 open FDs
> $ lsof -p 12154 | wc -l -> returned 531 FDs
> There were 228 threads running for the task manager.
>
> Drilling down a bit further, looking at a_inode and FIFO entries:
> $ lsof -p 12154 | grep a_inode | wc -l = 100 FDs
> $ lsof -p 12154 | grep FIFO | wc -l = 200 FDs
> $ /proc/12154/maps = 920 entries.
> Apart from lsof identifying lots of JARs and SOs being referenced there were
> also 244 child processes for the task manager process.
> Noticed that in each environment, a creep of file descriptors...are the above
> figures deemed excessive for the no of FDs in use? I know Flink uses Netty -
> is it using a separate Selector for reads & writes?
> Additionally Flink uses memory mapped files? or direct bytebuffers are these
> skewing the numbers of FDs shown?
> Example of one child process ID 6633:
> java 12154 6633 dfdev 387u a_inode 0,9 0 5869 [eventpoll]
> java 12154 6633 dfdev 388r FIFO 0,8 0t0 459758080 pipe
> java 12154 6633 dfdev 389w FIFO 0,8 0t0 459758080 pipe
> Lasty, cannot identify yet the reason for the creep in FDs even if Flink is
> pretty dormant or has dormant jobs. Production nodes are not experiencing
> excessive amounts of throughput yet either.
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
