> why sometime ,increase reducer number will not decrease job complete time ?
Apart from valid information that Yong wrote in the previous point, please note that: 1) You do not want to have very shortly lived (seconds) reduce tasks, because the overhead for coordinating them, starting JVMs, setting up the connections to all map tasks becomes too costly. It depends on your use case, but usually MapReduce jobs are for batch processing, and at my company we set the number of reduce tasks to make sure that each task runs at least a couple of minutes (for production jobs that are scheduled in "background", we aim for ~10 minutes). 2) We you have more reduce tasks, then you need more slots (or containers, if you use YARN). Sometimes, you can not get slots/containers as quick as you want, so that you can get stuck waiting for more resources. Then job completion time extends. 3) It you have thinner reducers, then they probably they write smaller output files to HDFS. Small files are problematic for HDFS (e.g. higher memory requirement on NN, bigger load on NN, slower NN restarts, more random than streaming access pattern and more). If the output of that job is later processed by another job, then you will see thin mappers (this can be partially alleviated by CombineFileInputFormat, though). 2013/12/11 java8964 <java8...@hotmail.com> > The whole job complete time depends on a lot of factors. Are you sure the > reducers part is the bottleneck? > > Also, it also depends on how many Reducer input groups it has in your MR > job. If you only have 20 reducer groups, even you jump your reducer count > to 40, then the epoch of reducers part won't have too much change, as the > additional 20 reducer task won't get data to process. > > If you have a lot of reducer input groups, and your cluster does have > capacity at this time, and your also have a lot idle reducer slot, then > increase your reducer count should decrease your whole job complete time. > > Make sense? > > Yong > > ------------------------------ > Date: Wed, 11 Dec 2013 14:20:24 +0800 > Subject: Re: issue about Shuffled Maps in MR job summary > From: justlo...@gmail.com > To: user@hadoop.apache.org > > > i read the doc, and find if i have 8 reducer ,a map task will output 8 > partition ,each partition will be send to a different reducer, so if i > increase reduce number ,the partition number increase ,but the volume on > network traffic is same,why sometime ,increase reducer number will not > decrease job complete time ? > > On Wed, Dec 11, 2013 at 1:48 PM, Vinayakumar B > <vinayakuma...@huawei.com>wrote: > > It looks simple, J > > > > Shuffled Maps= Number of Map Tasks * Number of Reducers > > > > Thanks and Regards, > > Vinayakumar B > > > > *From:* ch huang [mailto:justlo...@gmail.com] > *Sent:* 11 December 2013 10:56 > *To:* user@hadoop.apache.org > *Subject:* issue about Shuffled Maps in MR job summary > > > > hi,maillist: > > i run terasort with 16 reducers and 8 reducers,when i double > reducer number, the Shuffled maps is also double ,my question is the job > only run 20 map tasks (total input file is 10,and each file is 100M,my > block size is 64M,so split is 20) why i need shuffle 160 maps in 8 reducers > run and 320 maps in 16 reducers run?how to caculate the shuffle maps number? > > > > 16 reducer summary output: > > > > > > Shuffled Maps =320 > > > > 8 reducer summary output: > > > > Shuffled Maps =160 > > >
