[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16840122#comment-16840122
]
Ruiguang Pei commented on SPARK-24374:
--
Hi, [~mengxr]
For the map-reduce stage, the number of tasks(or data partitions) can bigger
than the number of total cores, so that the calculation cost of each data
partitions can be controlled, but when I'm using Barrier Execution Mode, it
seems that I mustn't re-partition my data more than number of total cores, if
you don't want to get "no enough resource" exception. Will it take some
computational performance problem? also, is this kind of design a little
inflexible?
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16766550#comment-16766550
]
Xiangrui Meng commented on SPARK-24374:
---
[~luzengxiang] When you create external processes, did you handle interrupted
exceptions or have a finally block to terminate the external process?
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16765694#comment-16765694
]
luzengxiang commented on SPARK-24374:
-
Hi [~mengxr], I am using Scala API.
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16739565#comment-16739565
]
Xiangrui Meng commented on SPARK-24374:
---
[~luzengxiang] Did you use Python API or Scala API for barrier mode? In
PySpark, Python worker process launches the external user process
(MPI/TensorFlow). However, the Python worker process might receive SIGKILL
directly from Python daemon. Hence we cannot rely on the task closure to clean
up (or we change how Spark terminates Python worker). A workaround is to watch
the parent Python worker process in the MPI job. If the parent is gone,
terminate itself. If you are on linux, you might use PR_SET_PDEATHSIG from
prctl.
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16738622#comment-16738622
]
Thomas Graves commented on SPARK-24374:
---
[~luzengxiang] are you just saying when spark tries to kill the tasks running
on tensorflow they don't really get killed? this could be tensorflow
spark.task.reaper.killTimeout.
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16737849#comment-16737849
]
luzengxiang commented on SPARK-24374:
-
I am tryging embedding MPI in barrier mode to support distributed tensoflow,
just like what said in this SPIP. The problem is MPI tasks are still runnig
when spark task is interrupt, which means by using barrier mode I can only
arise tensorflow training task but can not stop it. How can I work it out?
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-24374) SPIP: Support Barrier Execution Mode in Apache Spark
[
https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16728020#comment-16728020
]
Debasish Das commented on SPARK-24374:
--
Hi [~mengxr] with barrier mode available is it not possible to use native TF
parameter server in place of using MPI ? Although we are offloading compute
from spark to tf workers/ps, still if there is an exception that comes out,
tracking it with native TF API might be easier than MPI exception...great work
by the way...I was looking for a cloud-ml alternative using spark over
aws/azure/gcp and looks like barrier should help a lot although I am still not
clear on the limitations of TensorflowOnSpark project from Yahoo
[https://github.com/yahoo/TensorFlowOnSpark] which tried to put barrier like
syntax but not sure if few partitions fails on some tfrecord read /
communication exceptions whether it can re-run full job or it will only re-run
the failed partition...I guess the exception from few partitions can be thrown
back to spark driver and driver can take the action for re-run..when multiple
tf training jobs get scheduled on the same spark cluster I suspect TFoS might
have issues as well...
> SPIP: Support Barrier Execution Mode in Apache Spark
>
>
> Key: SPARK-24374
> URL: https://issues.apache.org/jira/browse/SPARK-24374
> Project: Spark
> Issue Type: Epic
> Components: ML, Spark Core
>Affects Versions: 2.4.0
>Reporter: Xiangrui Meng
>Assignee: Xiangrui Meng
>Priority: Major
> Labels: Hydrogen, SPIP
> Attachments: SPIP_ Support Barrier Scheduling in Apache Spark.pdf
>
>
> (See details in the linked/attached SPIP doc.)
> {quote}
> The proposal here is to add a new scheduling model to Apache Spark so users
> can properly embed distributed DL training as a Spark stage to simplify the
> distributed training workflow. For example, Horovod uses MPI to implement
> all-reduce to accelerate distributed TensorFlow training. The computation
> model is different from MapReduce used by Spark. In Spark, a task in a stage
> doesn’t depend on any other tasks in the same stage, and hence it can be
> scheduled independently. In MPI, all workers start at the same time and pass
> messages around. To embed this workload in Spark, we need to introduce a new
> scheduling model, tentatively named “barrier scheduling”, which launches
> tasks at the same time and provides users enough information and tooling to
> embed distributed DL training. Spark can also provide an extra layer of fault
> tolerance in case some tasks failed in the middle, where Spark would abort
> all tasks and restart the stage.
> {quote}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
