Thanks for the link. This is indeed what I am talking about.
The authors have taken the idea even further, avoiding disk writes on either the mapper or reducer side. It's not clear to me that this scales well to 1000s of nodes however, as the downside to not landing data on disk on the reducer side is that it would seem to impose at least one of the following requirements:
-- a lot of memory on the reducer side
-- reducers keep connections open to retrieve map file data
-- reducer/map-file connections are juggled so as to avoid keeping too many open at once.
From: Albert Chu [mailto:[EMAIL PROTECTED]]
Sent: Wednesday, June 12, 2013 2:27 PM
To: [EMAIL PROTECTED]
Subject: RE: Shuffle design: optimization tradeoffs
On Wed, 2013-06-12 at 18:08 +0000, John Lilley wrote:
> In reading this link as well as the sailfish report, it strikes me
> that Hadoop skipped a potentially significant optimization. Namely,
> why are multiple sorted spill files merged into a single output file?
> Why not have the auxiliary service merge on the fly, thus avoiding
> landing them to disk?
I believe what you're talking about/suggesting is similar to what's discussed in this paper?
> Was this considered and rejected due to placing memory/CPU
> requirements on the auxiliary service? I am assuming that whether the
> merge was done on disk or in a stream, it would require
> decompression/recompression of the data.
> -----Original Message-----
> From: Albert Chu [mailto:[EMAIL PROTECTED]]
> Sent: Tuesday, June 11, 2013 3:32 PM
> To: [EMAIL PROTECTED]
> Subject: Re: Shuffle design: optimization tradeoffs
> On Tue, 2013-06-11 at 16:00 +0000, John Lilley wrote:
> > I am curious about the tradeoffs that drove design of the
> > partition/sort/shuffle (Elephant book p 208). Doubtless this has
> > been tuned and measured and retuned, but I’d like to know what
> > observations came about during the iterative optimization process to
> > drive the final design. For example:
> > · Why does the mapper output create a single ordered file
> > containing all partitions, as opposed to a file per group of
> > partitions (which would seem to lend itself better to multi-core
> > scaling), or even a file per partition?
> I researched this awhile back wondering the same thing, and found this
> > · Why does the max number of streams to merge at once
> > (is.sort.factor) default to 10? Is this obsolete? In my
> > experience, so long as you have memory to buffer each input at 1MB
> > or so, the merger is more efficient as a single phase.
> > · Why does the mapper do a final merge of the spill files do
> > disk, instead of having the auxiliary process (in YARN) merge and
> > stream data on the fly?
> > · Why do mappers sort the tuples, as opposed to only
> > partitioning them and letting the reducers do the sorting?
> > Sorry if this is overly academic, but I’m sure a lot of people put a
> > lot of time into the tuning effort, and I hope they left a record of
> > their efforts.
> > Thanks
> > John
> Albert Chu
> [EMAIL PROTECTED]
> Computer Scientist
> High Performance Systems Division
> Lawrence Livermore National Laboratory
High Performance Systems Division
Lawrence Livermore National Laboratory
Bertrand Dechoux 2013-06-18, 22:04