On Thu, Jun 17, 2010 at 9:37 AM, Jeff Zhang <zjf...@gmail.com> wrote:
> Todd, > > Why's there a sorting in map task, the sorting here seems useless in my > opinion. > > For map-only jobs there isn't. For jobs with reduce, typically the number of reduce tasks is smaller than the number of map tasks, so parallelizing the sort on the mappers and just doing merge on the reducers is beneficial. Second, this allows the combiner to run on the mapper by identifying when it has multiple outputs for the same key. Third, this allows improved compression on the map output (thus less intermediate data transfer) by putting similar keys near each other (hopefully within the compression window). Fourth, it kills two birds with one stone since the mappers already have to group outputs by the partition. -Todd > > > On Thu, Jun 17, 2010 at 9:26 AM, Todd Lipcon <t...@cloudera.com> wrote: > > On Thu, Jun 17, 2010 at 12:43 AM, Jeff Zhang <zjf...@gmail.com> wrote: > > > >> Your understanding of Sort is not right. The key concept of Sort is > >> the TotalOrderPartitioner. Actually before the map-reduce job, client > >> side will do sampling of input data to estimate the distribution of > >> input data. And the mapper do nothing, each reducer will fetch its > >> data according the TotalOrderPartitioner. The data in each reducer is > >> local sorted, and each reducer are sorted ( r0<r1<r2....), so the > >> overall result data is sorted. > >> > > > > The sorting happens on the map side, actually, during the spill process. > The > > mapper itself is an identity function, but the map task code does perform > a > > sort (on a <partition,key> tuple) as originally described in this thread. > > Reducers just do a merge of mapper outputs. > > > > -Todd > > > > > >> > >> > >> > >> On Thu, Jun 17, 2010 at 12:13 AM, 李钰 <car...@gmail.com> wrote: > >> > Hi all, > >> > > >> > I'm doing some tuning of the sort benchmark of hadoop. To be more > >> specified, > >> > running test against the org.apache.hadoop.examples.Sort class. As > >> looking > >> > through the source code, I think the map tasks take responsibility of > >> > sorting the input data, and the reduce tasks just merge the map > outputs > >> and > >> > write them into HDFS. But here I've got a question I couldn't > understand: > >> > the time cost of the reduce phase of each reduce task, that is writing > >> data > >> > into HDFS, is different from each other. Since the input data and > >> operations > >> > of each reduce task is the same, what reason will cause the execution > >> time > >> > different? Is there anything wrong of my understanding? Does anybody > have > >> > any experience on this? Badly need your help, thanks. > >> > > >> > Best Regards, > >> > Carp > >> > > >> > >> > >> > >> -- > >> Best Regards > >> > >> Jeff Zhang > >> > > > > > > > > -- > > Todd Lipcon > > Software Engineer, Cloudera > > > > > > -- > Best Regards > > Jeff Zhang > -- Todd Lipcon Software Engineer, Cloudera
