spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sandy Ryza <sandy.r...@cloudera.com>
Subject Re: SPARK-streaming app running 10x slower on YARN vs STANDALONE cluster
Date Mon, 29 Dec 2014 18:15:14 GMT
Are you setting --num-executors to 8?

On Mon, Dec 29, 2014 at 10:13 AM, Mukesh Jha <me.mukesh.jha@gmail.com>
wrote:

> Sorry Sandy, The command is just for reference but I can confirm that
> there are 4 executors and a driver as shown in the spark UI page.
>
> Each of these machines is a 8 core box with ~15G of ram.
>
> On Mon, Dec 29, 2014 at 11:23 PM, Sandy Ryza <sandy.ryza@cloudera.com>
> wrote:
>
>> Hi Mukesh,
>>
>> Based on your spark-submit command, it looks like you're only running
>> with 2 executors on YARN.  Also, how many cores does each machine have?
>>
>> -Sandy
>>
>> On Mon, Dec 29, 2014 at 4:36 AM, Mukesh Jha <me.mukesh.jha@gmail.com>
>> wrote:
>>
>>> Hello Experts,
>>> I'm bench-marking Spark on YARN (
>>> https://spark.apache.org/docs/latest/running-on-yarn.html) vs a
>>> standalone spark cluster (
>>> https://spark.apache.org/docs/latest/spark-standalone.html).
>>> I have a standalone cluster with 3 executors, and a spark app running on
>>> yarn with 4 executors as shown below.
>>>
>>> The spark job running inside yarn is 10x slower than the one running on
>>> the standalone cluster (even though the yarn has more number of workers),
>>> also in both the case all the executors are in the same datacenter so there
>>> shouldn't be any latency. On YARN each 5sec batch is reading data from
>>> kafka and processing it in 5sec & on the standalone cluster each 5sec batch
>>> is getting processed in 0.4sec.
>>> Also, In YARN mode all the executors are not getting used up evenly as
>>> vm-13 & vm-14 are running most of the tasks whereas in the standalone mode
>>> all the executors are running the tasks.
>>>
>>> Do I need to set up some configuration to evenly distribute the tasks?
>>> Also do you have any pointers on the reasons the yarn job is 10x slower
>>> than the standalone job?
>>> Any suggestion is greatly appreciated, Thanks in advance.
>>>
>>> YARN(5 workers + driver)
>>> ========================
>>> Executor ID Address RDD Blocks Memory Used DU  AT FT CT TT TT Input ShuffleRead
>>> ShuffleWrite Thread Dump
>>> 1 vm-18.cloud.com:51796 0 0.0B/530.3MB 0.0 B 1 0 16 17 634 ms 0.0 B 2047.0
>>> B 1710.0 B Thread Dump
>>> 2 vm-13.cloud.com:57264 0 0.0B/530.3MB 0.0 B 0 0 1427 1427 5.5 m 0.0 B 0.0
>>> B 0.0 B Thread Dump
>>> 3 vm-14.cloud.com:54570 0 0.0B/530.3MB 0.0 B 0 0 1379 1379 5.2 m 0.0 B 1368.0
>>> B 2.8 KB Thread Dump
>>> 4 vm-11.cloud.com:56201 0 0.0B/530.3MB 0.0 B 0 0 10 10 625 ms 0.0 B 1368.0
>>> B 1026.0 B Thread Dump
>>> 5 vm-5.cloud.com:42958 0 0.0B/530.3MB 0.0 B 0 0 22 22 632 ms 0.0 B 1881.0
>>> B 2.8 KB Thread Dump
>>> <driver> vm.cloud.com:51847 0 0.0B/530.0MB 0.0 B 0 0 0 0 0 ms 0.0 B 0.0
>>> B 0.0 B Thread Dump
>>>
>>> /homext/spark/bin/spark-submit
>>> --master yarn-cluster --num-executors 2 --driver-memory 512m
>>> --executor-memory 512m --executor-cores 2
>>> --class com.oracle.ci.CmsgK2H /homext/lib/MJ-ci-k2h.jar
>>> vm.cloud.com:2181/kafka spark-yarn avro 1 5000
>>>
>>> STANDALONE(3 workers + driver)
>>> ==============================
>>> Executor ID Address RDD Blocks Memory Used DU AT FT CT TT TT Input ShuffleRead
>>> ShuffleWrite Thread Dump
>>> 0 vm-71.cloud.com:55912 0 0.0B/265.0MB 0.0 B 0 0 1069 1069 6.0 m 0.0 B 1534.0
>>> B 3.0 KB Thread Dump
>>> 1 vm-72.cloud.com:40897 0 0.0B/265.0MB 0.0 B 0 0 1057 1057 5.9 m 0.0 B 1368.0
>>> B 4.0 KB Thread Dump
>>> 2 vm-73.cloud.com:37621 0 0.0B/265.0MB 0.0 B 1 0 1059 1060 5.9 m 0.0 B 2.0
>>> KB 1368.0 B Thread Dump
>>> <driver> vm.cloud.com:58299 0 0.0B/265.0MB 0.0 B 0 0 0 0 0 ms 0.0 B 0.0
>>> B 0.0 B Thread Dump
>>>
>>> /homext/spark/bin/spark-submit
>>> --master spark://chsnmvproc71vm3.usdc2.oraclecloud.com:7077
>>> --class com.oracle.ci.CmsgK2H /homext/lib/MJ-ci-k2h.jar
>>> vm.cloud.com:2181/kafka spark-standalone avro 1 5000
>>>
>>> PS: I did go through the spark website and
>>> http://www.virdata.com/tuning-spark/, but was out of any luck.
>>>
>>> --
>>> Cheers,
>>> Mukesh Jha
>>>
>>
>>
>
>
> --
>
>
> Thanks & Regards,
>
> *Mukesh Jha <me.mukesh.jha@gmail.com>*
>

Mime
View raw message