Hi,
I guess I did not frame my question properly. What was actually meant
was this: After the map phase, the output of each map is partitioned
based on the key value and written to disk as a single file. Now, the
ReducerTask starts up and has to read the intermediate values from the
partitions, that have been created for it by the various map tasks. Now,
how does the ReduceTask decide which partition to read first?
Thanks,
Virajith
On 6/29/2011 11:37 PM, David Rosenstrauch wrote:
On 06/29/2011 05:28 PM, Virajith Jalaparti wrote:
Hi,
I was wondering what scheduling algorithm is used in Hadoop (version
0.20.2 in particular), for a ReduceTask to determine in what order it is
supposed to read the map outputs from the various mappers that have been
run? In particular, suppose we have 10maps called map1, map2,....,
map10. and say 2 reducers r1 and r2. Which map's output does r1/r2 read
from first?
Also, suppose that the mapred.reduce.parallel.copies is set to 5. Then
do both r1 and r2 read from 5 map outputs concurrently?
Thanks,
Virajith
You're missing 2 key steps in here. After the mappers, a sort step
gets run (to sort the records in key order) and then a partition step
(to partition the records by key and spread them across the reducers).
So your question is really a moot one. The records output by a given
map step get spread across multiple reducers, and not all sent to a
single reducer.
DR
