spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gourav Sengupta <gourav.sengu...@gmail.com>
Subject Re: High Disk Usage In Spark 2.2.1 With No Shuffle Or Spill To Disk
Date Sun, 08 Apr 2018 03:51:21 GMT
Hi Saad,

May I ask which EMR version and cluster size you are using? Usually if you
are using C4.4xlarge systems then they have high disk space as I
understand.

The other thing that you can do is attach more disk space to the nodes, the
option of which is available in the advanced cluster start option, and then
change the cluster configuration LOCAL_DIRS to point to this new location
and also the spark.local.dir configuration to point to this location
(although this may be redundant).

I prefer to use auto scaling and spot instances for these kind of migration
activities.

Best of luck.


Regards,
Gourav

On Sat, Apr 7, 2018 at 10:09 PM, Saad Mufti <saad.mufti@gmail.com> wrote:

> I have been trying to monitor this while the job is running, I think I
> forgot to account for the 3-way hdfs replication, so right there the output
> is more like 21 TB instead of my claimed 7 TB. But it still looks like hdfs
> is losing more disk space than can be account for by just the output, going
> by the output of the dfsadmin command, so I am still trying to track that
> down. The total allocated disk space of 28 TB should still be more than
> enough.
>
> ----
> Saad
>
>
> On Sat, Apr 7, 2018 at 2:40 PM, Saad Mufti <saad.mufti@gmail.com> wrote:
>
>> Thanks. I checked and it is using another s3 folder for the temporary
>> restore space. The underlying code insists on the snapshot and the restore
>> directory being on the same filesystem, so it is using Emrfs for both. So
>> unless Emrfs is under the covers using some local disk space it doesn't
>> seem like that is responsible.
>>
>> ----
>> Saad
>>
>> On Sat, Apr 7, 2018 at 2:37 PM Jörn Franke <jornfranke@gmail.com> wrote:
>>
>>> As far as I know the TableSnapshotInputFormat relies on a temporary
>>> folder
>>> https://hbase.apache.org/apidocs/org/apache/hadoop/hbase/
>>> mapreduce/TableSnapshotInputFormat.html
>>>
>>>
>>> Unfortunately some inputformats need a (local) tmp Directory. Sometimes
>>> this cannot be avoided.
>>>
>>> See also the source:
>>> https://github.com/apache/hbase/blob/master/hbase-mapreduce/
>>> src/main/java/org/apache/hadoop/hbase/mapred/TableSnaps
>>> hotInputFormat.java
>>>
>>> On 7. Apr 2018, at 20:26, Saad Mufti <saad.mufti@gmail.com> wrote:
>>>
>>> Hi,
>>>
>>> I have a simple ETL Spark job running on AWS EMR with Spark 2.2.1 . The
>>> input data is HBase files in AWS S3 using EMRFS, but there is no HBase
>>> running on the Spark cluster itself. It is restoring the HBase snapshot
>>> into files on disk in another S3 folder used for temporary storage, then
>>> creating an RDD over those files using HBase's TableSnapsotInputFormat
>>> class. There is a large number of HBase regions, around 12000, and each
>>> region gets translated to one Spark task/partition. We are running in YARN
>>> mode, with one core per executor, so on our 120 node cluster we have around
>>> 1680 executors running (not the full 1960 as YARN only gives us so many
>>> containers due to memory limits).
>>>
>>> This is a simple ETL job that transforms the HBase data into
>>> Avro/Parquet and writes to disk, there are no reduces or joins of any kind.
>>> The output Parquet data is using Snappy compression, the total output is
>>> around 7 TB while we have about 28 TB total disk provisioned in the
>>> cluster. The Spark UI shows no disk storage being used for cached data, and
>>> not much heap being used for caching either, which makes sense because in
>>> this simple job we have no need to do RDD.cache as the RDD is not reused at
>>> all.
>>>
>>> So lately the job has started failing because close to finishing, some
>>> of the YARN nodes start running low on disk and YARN marks them as
>>> unhealthy, then kills all the executors on that node. But the problem just
>>> moves to another node where the tasks are relaunched for another attempt
>>> until after 4 failures for a given task the whole job fails.
>>>
>>> So I am trying to understand where all this disk usage is coming from? I
>>> can see in Ganglia that disk is running low the longer the job runs no
>>> matter which node I look at. Like I said the total output size of the final
>>> output in hdfs is only around 7 TB while we have around 28 TB of disk
>>> provisioned for hdfs.
>>>
>>> Any advice or pointers for where to look for the large disk usage would
>>> be most appreciated.
>>>
>>> Thanks.
>>>
>>> ----
>>> Saad
>>>
>>>
>

Mime
View raw message