The approach which I proposed will have m+n i/o for reading datasets not the (m + n + m*n) and but further i/o due to spills and reading mapper output by reducer will be more as number of tuples coming out of mapper are ( m + m * n).
On 18-Apr-2013, at 5:40 PM, zheyi rong wrote:
Thank you, now I get your point.
But I wonder that this approach would be slower than
implementing a custom InputFormat which, each time, provides a pair of lines to mappers; then doing the product in mappers? (in
Since your approach would need (m + n + m*n) I/O in mapper side, and (2*m*n) IO in reducer side;
while with implementing a custom InputFormat, the I/O is (m*n).
I am asking this because I have implemented the custom InputFormat, but the running time is still intolerable in our small cluster.
On Thu, Apr 18, 2013 at 1:45 PM, Ajay Srivastava <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:
Yes, that's a crucial part.
Write a class which extends WritableComparator and override compare method.
You need to set this class in job client as -
job.setGroupingComparatorClass (Grouping comparator class).
This will make sure that records having same Ki will be grouped together and will go to same iteration of reduce.
I forgot to mention in my previous post to write a partitioner too which partitions data based on first part of key.
On 18-Apr-2013, at 4:42 PM, zheyi rong wrote:
Hi Ajay Srivastava,
Thank your for your reply.
Could you please explain a little bit more on "Write a grouping comparator which group records on first part of key i.e. Ki." ?
I guess it is a crucial part, which could filter some pairs before passing them to the reducer.
On Thu, Apr 18, 2013 at 12:50 PM, Ajay Srivastava <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:
You can use following simple method.
Lets say dataset1 has m records and when you emit these records from mapper, keys are K1,K2 ….., Km for each respective record. Also add an identifier to identify dataset from where records is being emitted.
So if R1 is a record in dataset1, the mapper will emit key as (K1, DATASET1) and value as R1.
For dataset2 having n records, emit m records for each record with keys K1, K2, …., Km and identifier as DATASET2.
So if R1' is a record from dataset2, emit m records with key as (Ki, DATASET2) and value R1' where i is from 1 to m.
Write a grouping comparator which group records on first part of key i.e. Ki.
In reducer, for each iteration of reduce there will be one record from dataset1 and n records from dataset2. Get the cartesian product, apply filter and then output.
Note -- You may not know keys (K1, K2, … , Km) before hand. If yes, then you need one more pass of dataset1 to identify the keys and store it to use for dataset2.
On 18-Apr-2013, at 3:51 PM, Azuryy Yu wrote:
This is not suitable for his large dataset.
--Send from my Sony mobile.
On Apr 18, 2013 5:58 PM, "Jagat Singh" <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:
Can you have a look at
On Thu, Apr 18, 2013 at 7:47 PM, zheyi rong <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:
I am writing to kindly ask for ideas of doing cartesian product in hadoop.
Specifically, now I have two datasets, each of which contains 20million lines.
I want to do cartesian product on these two datasets, comparing lines pairwisely.
The output of each comparison can be mostly filtered by a function ( we do not output the
whole result of this cartesian product, but only a small part).
I guess one good way is to pass one block from dataset1 and another block from dataset2
to a mapper, then let the mappers do the product in memory to avoid IO.
Thank you very much.