[ https://issues.apache.org/jira/browse/SPARK-24374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16498465#comment-16498465 ]
Henry Robinson commented on SPARK-24374: ---------------------------------------- The use case in the SPIP isn't 100% convincing. I'm concerned about the idea of embedding other execution engines inside Spark tasks, effectively using Spark's job scheduler as a cluster manager. Resource consumption would be very different at different phases of the specified task (some tasks would be waiting, one task would have launched MPI) making it hard for sensible allocation decisions to be made. Do you anticipate requiring support from the cluster manager for gang scheduling? Or is {{barrier()}} going to be enough to ensure that all tasks wait - even if they take up an executor slot on the cluster for a long time waiting for the other tasks to get scheduled? An alternative design for the example in the SPIP document would be to split it into more than one Spark job, e.g.: # (Spark job) write the input files in parallel # (not via Spark) launch the MPI job via the cluster manager (perhaps a kubernetes pod, for example) # (Spark job) consume the output files in parallel Spark already has semantics that allow you to wait until a job is finished so the driver can naturally coordinate amongst the different phases without needing to add extra coordination primitives to individual tasks. But synchronous execution is a thing in distributed systems, of course, so maybe there are more compelling use cases than the one in the SPIP? > SPIP: Support Barrier Scheduling 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: 3.0.0 > Reporter: Xiangrui Meng > Assignee: Xiangrui Meng > Priority: Major > Labels: 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