Can you show us the program that you are running. If you are setting number of partitions in the XYZ-ByKey operation as 300, then there should be 300 tasks for that stage, distributed on the 50 executors are allocated to your context. However the data distribution may be skewed in which case, you can use a repartition operation to redistributed the data more evenly (both DStream and RDD have repartition).
TD On Fri, Jul 11, 2014 at 12:22 AM, Bill Jay <bill.jaypeter...@gmail.com> wrote: > Hi Tathagata, > > I also tried to use the number of partitions as parameters to the > functions such as groupByKey. It seems the numbers of executors is around > 50 instead of 300, which is the number of the executors I specified in > submission script. Moreover, the running time of different executors is > skewed. The ideal case is that Spark can distribute the data into 300 > executors evenly so that the computation can be efficiently finished. I am > not sure how to achieve this. > > Thanks! > > Bill > > > On Thu, Jul 10, 2014 at 5:59 PM, Tathagata Das < > tathagata.das1...@gmail.com> wrote: > >> Can you try setting the number-of-partitions in all the shuffle-based >> DStream operations, explicitly. It may be the case that the default >> parallelism (that is, spark.default.parallelism) is probably not being >> respected. >> >> Regarding the unusual delay, I would look at the task details of that >> stage in the Spark web ui. It will show break of time for each task, >> including GC times, etc. That might give some indication. >> >> TD >> >> >> On Thu, Jul 10, 2014 at 5:13 PM, Bill Jay <bill.jaypeter...@gmail.com> >> wrote: >> >>> Hi Tathagata, >>> >>> I set default parallelism as 300 in my configuration file. Sometimes >>> there are more executors in a job. However, it is still slow. And I further >>> observed that most executors take less than 20 seconds but two of them take >>> much longer such as 2 minutes. The data size is very small (less than 480k >>> lines with only 4 fields). I am not sure why the group by operation takes >>> more then 3 minutes. Thanks! >>> >>> Bill >>> >>> >>> On Thu, Jul 10, 2014 at 4:28 PM, Tathagata Das < >>> tathagata.das1...@gmail.com> wrote: >>> >>>> Are you specifying the number of reducers in all the DStream.****ByKey >>>> operations? If the reduce by key is not set, then the number of reducers >>>> used in the stages can keep changing across batches. >>>> >>>> TD >>>> >>>> >>>> On Wed, Jul 9, 2014 at 4:05 PM, Bill Jay <bill.jaypeter...@gmail.com> >>>> wrote: >>>> >>>>> Hi all, >>>>> >>>>> I have a Spark streaming job running on yarn. It consume data from >>>>> Kafka and group the data by a certain field. The data size is 480k lines >>>>> per minute where the batch size is 1 minute. >>>>> >>>>> For some batches, the program sometimes take more than 3 minute to >>>>> finish the groupBy operation, which seems slow to me. I allocated 300 >>>>> workers and specify 300 as the partition number for groupby. When I >>>>> checked >>>>> the slow stage *"combineByKey at ShuffledDStream.scala:42",* there >>>>> are sometimes 2 executors allocated for this stage. However, during other >>>>> batches, the executors can be several hundred for the same stage, which >>>>> means the number of executors for the same operations change. >>>>> >>>>> Does anyone know how Spark allocate the number of executors for >>>>> different stages and how to increase the efficiency for task? Thanks! >>>>> >>>>> Bill >>>>> >>>> >>>> >>> >> >
