1) I am varying both the number of mappers and reducers trying to
determine three things
a) What are the options I need reducers and mappers to
- Not have mappers or reducers killed with GC overhead limit
exceeded
- Minimize execution time for the cluster
I use a custom Splitter and can adjust the block size for anywhere from 1
Mapper to hundreds of mappers -
For an 8 node cluster I am trying 8,16 and 24 reducers
2) I have been playing with
io,sort,factor - using 100 for now
io.sort.mb is 400 - use mich higher values and the job will not run
3) I set child,vm.opts to Xmx3000m ( getting simulat results to using 1300
4) My mappreads a single file about 1 GB in size. Each item the splitter
delivers (about 1KB) generates tens of thousands of Key Value pair
(<100bytes per value) I can do all the work of generating the output on one
machine (but not the shuffle and sort) in about an hour on one box but my
job is running
for a many hours without completing. I also got a lot of
after seeing in other tasks
Lost task tracker:
tracker_glados4.systemsbiology.net:localhost.localdomain/127.0.0.1:32790
Caused by: java.lang.NullPointerException
at
org.apache.hadoop.io.compress.BZip2Codec$BZip2CompressionOutputStream.write(BZip2Codec.java:200)
at
org.apache.hadoop.fs.FSDataOutputStream$PositionCache.write(FSDataOutputStream.java:41)
at java.io.DataOutputStream.writeByte(DataOutputStream.java:136)
at org.apache.hadoop.io.WritableUtils.writeVLong(WritableUtils.java:263)
at org.apache.hadoop.io.WritableUtils.writeVInt(WritableUtils.java:243)
at org.apache.hadoop.mapred.IFile$Writer.close(IFile.java:126)
at
org.apache.hadoop.mapred.MapTask$MapOutputBuffer.sortAndSpill(MapTask.java:1242)
at
org.apache.hadoop.mapred.MapTask$MapOutputBuffer.access$1800(MapTask.java:648)
at
org.apache.hadoop.mapred.MapTask$MapOutputBuffer$SpillThread.run(MapTask.java:1135)
The same job runs well with a smaller data set. Most of the reason for
moving to hadoop is to allow solutions to scale and I am very concerned at
how
badly my larger cases are doing. The documentation does not say about about
tuning parameters for my larger jobs without running into swap hell or
GC overhead limit exceeeded.
On Sun, Nov 6, 2011 at 7:11 AM, Florin P <florinp...@yahoo.com> wrote:
> Hello!
>
> How many reducers you are using?
> Regarding the performance parameters, fist you can increase the size of
> the io.sort.mb parameter.
> It seems that you are sending a lot of amount of data to the reducer. By
> increasing the value of this parameter, in the shuffle phase, the framework
> will not be forced to write/spill data on the HDD that could be a reason
> for slowing the process.
> If you are using one reducer, then the whole data is sent over HTTP to
> that reducer. Another thing that you have to think about it.
> Just for a curiosity, try increase also the dfs.block.size to 128 MB. It
> seems that you are using the default 64 MB. You'll get less mapper tasks.
> Also, depending what configuration you have on the machine how many cores
> do you have on CPU, you can increase the values for
> mapred.tasktracker.{map|reduce}.tasks.maximum The maximum number of
> Map/Reduce tasks, which are run simultaneously on a given TaskTracker,
> individually. Defaults to 2 (2 maps and 2 reduces), but vary it
> depending on your hardware
> You can have a look at
> http://hadoop.apache.org/common/docs/r0.20.2/cluster_setup.html.
> A good book for understanding tuning parameters is Hadoop Definitive
> Guide by Tom White.
>
> Hope that the above helps.
> Regards,
> Florin
>
>
>
>
> --- On Thu, 11/3/11, Steve Lewis <lordjoe2...@gmail.com> wrote:
>
> From: Steve Lewis <lordjoe2...@gmail.com>
> Subject: Problems with MR Job running really slowly
> To: "mapreduce-user" <mapreduce-user@hadoop.apache.org>
> Date: Thursday, November 3, 2011, 11:07 PM
>
> I have a job which takes an xml file - the splitter breaks the file into
> tags, the mapper parses each tag and sends the data to the reducer. I am
> using a custom splitter which reads
> the file looking for start and end tags.
>
>
> When I run the code in the splitter and the mapper - generating separate
> tags and parsing them I can read a file sized at about 500MB containing
> 12000 tags on my local system in 23 seconds
>
>
> When I read a file on HDFS on a local cluster I can read and parse the
> file in 38 seconds
> When I run the same code on a eight node cluster I get 7 map tasks. The
> mappers are taking 190 seconds to handle 100 tags of
>
> which 200 millisec is parsing and almost all of the rest of the time is
> in context.write. A mapper handling 1600 tags takes about 3 hours -These
> are the statistics for a map task - it it true that one tag well be sent to
> about 300 keys but still 3 hours to write 1,5 million records and 5Gb seems
> way excessive
> FileSystemCountersFILE_BYTES_READ 816,935,457HDFS_BYTES_READ
> 439,554,860FILE_BYTES_WRITTEN 1,667,745,197
> PerformanceTotalScoredScans 1,660
>
> Map-Reduce FrameworkCombine output records0Map input records 6,134Spilled
> Records 1,690,063Map output bytes 5,517,423,780Combine input records 0
>
>
> Map output records 571,475
> Anyone want to offer suggestions on how to tune the job better
>
> --
> Steven M. Lewis PhD4221 105th Ave NEKirkland, WA 98033
>
> 206-384-1340 (cell)
> Skype lordjoe_com
>
>
>
>
>
>
--
Steven M. Lewis PhD
4221 105th Ave NE
Kirkland, WA 98033
206-384-1340 (cell)
Skype lordjoe_com
