> On 2011-12-12 02:10:01, Dmitriy Lyubimov wrote: > > Hm. I hope i did not read the code or miss something. > > > > 1 -- i am not sure this will actually work as intended unless # of reducers > > is corced to 1, of which i see no mention in the code. > > 2 -- mappers do nothing, passing on all the row pressure to sort which is > > absolutely not necessary. Even if you use combiners. This is going to be > > especially the case if you coerce 1 reducer an no combiners. IMO mean > > computation should be pushed up to mappers to avoid sort pressures of map > > reduce. Then reduction becomes largely symbolical(but you do need pass on > > the # of rows mapper has seen, to the reducer, in order for that operation > > to apply correctly). > > 3 -- i am not sure -- is NullWritable as a key legit? In my experience > > sequence file reader cannot instantiate it because NullWritable is a > > singleton and its creation is prohibited by making constructor private. > > Raphael Cendrillon wrote: > Thanks Dmitry. > > Regarding 1, if I understand correctly the number of reducers depends on > the number of unique keys. Since all keys are set to the same value (null), > then all of the mapper outputs should arrive at the same reducer. This seems > to work in the unit test, but I may be missing something? > > Regarding 2, that makes alot of sense. I'm wondering how many rows should > be processed per mapper? I guess there is a trade-off between scalability > (processing more rows within a single map job means that each row must have > less columns) and speed? Is there someplace in the SSVD code where the > matrix is split into slices of rows that I could use as a reference? > > Regarding 3, I believe NullWritable is OK. It's used pretty extensively > in TimesSquaredJob in DistributedRowMatrx. However if you feel there is some > disadvantage to this I could replace "NullWritable.get()" with "new > IntWritable(1)" (that is, set all of the keys to 1). Would that be more > suitable? > > > > Dmitriy Lyubimov wrote: > NullWritable objection is withdrawn. Apparently i haven't looked into > hadoop for too long, amazingly it seems to work now. > > > 1 -- I don't think your statement about # of reduce tasks is true. > > The job (or, rather, user) sets the number of reduce tasks via config > propery. All users will follow hadoop recommendation to set that to 95% of > capacity they want to take. (usually the whole cluster). So in production > environment you are virtually _guaranteed_ to have number of reducers of > something like 75 on a 40-noder and consequently 75 output files (unless > users really want to read the details of your job and figure you meant it to > be just 1). > Now, it is true that only one file will actually end up having something > and the rest of task slots will just be occupied doing nothing . > > So there are two problems with that scheme: a) is that job that allocates > so many task slots that do nothing is not a good citizen, since in real > production cluster is always shared with multiple jobs. b) your code assumes > result will end up in partition 0, whereas contractually it may end up in any > of 75 files. (in reality with default hash partitioner for key 1 it will wind > up in partion 0001 unless there's one reducer as i guess in your test was). > > 2-- it is simple. when you send n rows to reducers, they are shuffled - > and - sorted. Sending massive sets to reducers has 2 effects: first, even if > they all group under the same key, they are still sorted with ~ n log (n/p) > where p is number of partitions assuming uniform distribution (which it is > not because you are sending everything to the same place). Just because we > can run distributed sort, doesn't mean we should. Secondly, all these rows > are physically moved to reduce tasks, which is still ~n rows. Finally what > has made your case especially problematic is that you are sending everything > to the same reducer, i.e. you are not actually doing sort in distributed way > but rather simple single threaded sort at the reducer that happens to get all > the input. > > So that would allocate a lot of tasks slots that are not used; but do a > sort that is not needed; and do it in a single reducer thread for the entire > input which is not parallel at all. > > Instead, consider this: map has a state consisting of (sum(X), k). it > keeps updating it sum+=x, k++ for every new x. At the end of the cycle (in > cleanup) it writes only 1 tuple (sum(x), k) as output. so we just reduced > complexity of the sort and io from millions of elements to just # of maps > (which is perhaps just handful and in reality rarely overshoots 500 mappers). > That is, about at least 4 orders of magnitude. > > Now, we send that handful tuples to single reducer and just do combining > (sum(X)+= sum_i(X); n+= n_i) where i is the tuple in reducer. And because it > is only a handful, reducer also runs very quickly, so the fact that we > coerced it to be 1, is pretty benign. That volume of anywhere between 1 to > 500 vectors it sums up doesn't warrant distributed computation. > > But, you have to make sure there's only 1 reducer no matter what user put > into the config, and you have to make sure you do all heavy lifting in the > mappers. > > Finally, you don't even to coerce to 1 reducer. You still can have > several (but uniformly distributed) and do final combine in front end of the > method. However, given small size and triviality of the reduction processing, > it is probably not warranted. Coercing to 1 reducer is ok in this case IMO. > > 3 i guess any writable is ok but NullWritable. Maybe something has > changed. i remember falling into that pitfall several generations of hadoop > ago. You can verify by staging a simple experiment of writing a sequence file > with nullwritable as either key or value and try to read it back. in my test > long ago it would write ok but not read back. I beleive similar approach is > used with keys in shuffle and sort. There is a reflection writable factory > inside which is trying to use default constructor of the class to bring it up > which is(was) not available for NullWritable. > > >
Thanks Dmitriy. I've updated the diff to push the row summation into the mapper as you suggested, force the number of reducers to 1, and make the final output key IntWritable. Could you please take a look? - Raphael ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/3147/#review3838 ----------------------------------------------------------- On 2011-12-12 10:41:46, Raphael Cendrillon wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/3147/ > ----------------------------------------------------------- > > (Updated 2011-12-12 10:41:46) > > > Review request for mahout. > > > Summary > ------- > > Here's a patch with a simple job to calculate the row mean (column-wise > mean). One outstanding issue is the combiner, this requires a wrtiable class > IntVectorTupleWritable, where the Int stores the number of rows, and the > Vector stores the column-wise sum. > > > This addresses bug MAHOUT-923. > https://issues.apache.org/jira/browse/MAHOUT-923 > > > Diffs > ----- > > > /trunk/core/src/main/java/org/apache/mahout/math/hadoop/DistributedRowMatrix.java > 1213095 > > /trunk/core/src/main/java/org/apache/mahout/math/hadoop/MatrixRowMeanJob.java > PRE-CREATION > > /trunk/core/src/test/java/org/apache/mahout/math/hadoop/TestDistributedRowMatrix.java > 1213095 > > Diff: https://reviews.apache.org/r/3147/diff > > > Testing > ------- > > Junit test > > > Thanks, > > Raphael > >
