Re: Map Shuffle Bytes

[Eduard Skaley]

For this I need to know where an inputsplit is located. And where a join 
is computed. How can I do this programmatically ?
This isn't called 'shuffle' (but rather a plain remote read) so your
original question was confusing, thanks for clarifying!
In that case, you could count the bytes coming in from the required
record reader - for example a TextRecordReader uses a Long key that
denotes current offset in file, which you could use as a simple,
progressing counter of bytes read thus far.

On Wed, Dec 26, 2012 at 5:16 PM, Eduard Skaley <e.v.ska...@gmail.com> wrote:
Hi,

I mean TO the mappers. I'm using the CompositeInputFormat for my application
to compute map-side joins.
I want to join two datasets A and B one is stored on node 1 and the other
one on node 2.
For example if the join will be computed on node 2 then the inputsplit of
the dataset which is stored on node 1 has to be transferred to node 2.
I want to count the bytes which are shuffled (transferred) TO the mapper of
node 2.

Hi,

What do you mean by "shuffled bytes [to] the mappers"? If you mean
"from", it is "Reduce shuffle bytes" you look for; otherwise, you may
be looking for the per-map counter of "Map output bytes".

Per-partition counters can be constructed on the user side if needed,
by pre-computing the partition before emit (using the same
partitioner) and counting up the bytes of your objects for its
counter.

On Tue, Dec 25, 2012 at 6:03 PM, Eduard Skaley <e.v.ska...@gmail.com>
wrote:
Hello guys,

I need a counter for shuffled bytes to the mappers.
Is there existing one or should I define one myself ?
How can I implement such a counter?

Thank you and happy Christmas time,
Eduard