Issues with Job running on multiple nodes

sshettar · Post by **sshettar** » Fri Aug 28, 2009 10:18 am

Hi All,

There is this job which is curently running on 1 node , i made few changes to this job and ran the job on 1 node itself and its working fine , now i wanted to check the performance running on 4 nodes .
the job ran successfully but when i checked the dataset ( which is what i load at the end of my job) , i see that the data is flowing only through 1 node and the rest 3 nodes have 0 records through them .

ANy light on why this could be happening?

Any help is highly appraciated.

Thanks
Shalini

chulett · Post by **chulett** » Fri Aug 28, 2009 12:50 pm

What kind of partitioning are you doing in the job?

sshettar · Post by **sshettar** » Fri Aug 28, 2009 12:52 pm

I have set the partitioning to Auto itself

chulett · Post by **chulett** » Fri Aug 28, 2009 12:55 pm

OK, how about some details of your job design? Anything running in Sequential rather than Parallel mode?

Sainath.Srinivasan · Post by **Sainath.Srinivasan** » Fri Aug 28, 2009 1:01 pm

Is the source a sequential file or very small volume ?

sshettar · Post by **sshettar** » Fri Aug 28, 2009 1:14 pm

The Job is doing several joins. ( all the joins are necessary) and the data is sorted within the database stages itself using the order by clause. but the partitioing i have left it to auto since there were several joins and i thought specifying the partitioing and sorting would probably degrae the performance. hence used auto instead. finally a dataset gets loaded.

Join Stage1---5 DBStages---output1
output1 joined with Dbsatge ---output2
....
...
...

similarly after 6 such joins the data is passed through the transformer and then loads the dataset.

Hope i am fairly clear in explaining with my design

Thanks in advance

Raftsman · Post by **Raftsman** » Fri Aug 28, 2009 1:57 pm

If your volume is small, only on node will be used. Increasing nodes doesn't necessarily mean better throughput. It may take longer for the job to setup with 4 nodes. Add both the startup and CPU time to determine which is the optimal setting. I have found that one node is more efficient most of the time.

miwinter · Post by **miwinter** » Tue Sep 01, 2009 1:39 am

How many records are being processed through this job (including on each side of each join)?

Add environment variable $APT_DUMP_SCORE with value True and rerun the job, then copy and paste the output from the job log here please.