Hi Alan, Cheolsoo, I re-ran the benchmarks with and without the combiner. Enabling the combiner is faster:
With combiner: real 668.44 real 663.10 real 665.05 Without combiner: real 795.97 real 810.51 real 810.16 Best Regards, Ben On 22 August 2013 16:33, Cheolsoo Park <[email protected]> wrote: > Hi Benjamin, > > To answer your question, how the Hadoop combiner works is that 1) mappers > write outputs to disk and 2) combiners read them, combine and write them > again. So you're paying extra disk I/O as well as > serialization/deserialization. > > This will pay off if combiners significantly reduce the intermediate > outputs that reducers need to fetch from mappers. But if reduction is not > significant, it will only slow down mappers. You can identify whether this > is really a problem by comparing the time spent by map and combine > functions in the task logs. > > What I usually do are: > 1) If there are many small bags, disable combiners. > 2) If there are many large bags, enable combiners. Furthermore, turning on > "pig.exec.mapPartAgg" helps. (see the Pig > blog<https://blogs.apache.org/pig/entry/apache_pig_it_goes_to>for > details. > ) > > Thanks, > Cheolsoo > > > On Thu, Aug 22, 2013 at 4:01 AM, Benjamin Jakobus <[email protected] > >wrote: > > > Hi Cheolsoo, > > > > Thanks - I will try this now and get back to you. > > > > Out of interest; could you explain (or point me towards resources that > > would) why the combiner would be a problem? > > > > Also, could the fact that Pig builds an intermediary data structure (?) > > whilst Hive just performs a sort then the arithmetic operation explain > the > > slowdown? > > > > (Apologies, I'm quite new to Pig/Hive - just my guesses). > > > > Regards, > > Benjamin > > > > > > On 22 August 2013 01:07, Cheolsoo Park <[email protected]> wrote: > > > > > Hi Benjamin, > > > > > > Thank you very much for sharing detailed information! > > > > > > 1) From the runtime numbers that you provided, the mappers are very > slow. > > > > > > CPU time spent (ms)5,081,610168,7405,250,350CPU time spent > (ms)5,052,700 > > > 178,2205,230,920CPU time spent (ms)5,084,430193,4805,277,910 > > > > > > 2) In your GROUP BY query, you have an algebraic UDF "COUNT". > > > > > > I am wondering whether disabling combiner will help here. I have seen a > > lot > > > of cases where combiner actually hurt performance significantly if it > > > doesn't combine mapper outputs significantly. Briefly looking at > > > generate_data.pl in PIG-200, it looks like a lot of random keys are > > > generated. So I guess you will end up with a large number of small bags > > > rather than a small number of large bags. If that's the case, combiner > > will > > > only add overhead to mappers. > > > > > > Can you try to include this "set pig.exec.nocombiner true;" and see > > whether > > > it helps? > > > > > > Thanks, > > > Cheolsoo > > > > > > > > > > > > > > > > > > > > > On Wed, Aug 21, 2013 at 3:52 AM, Benjamin Jakobus < > > [email protected] > > > >wrote: > > > > > > > Hi Cheolsoo, > > > > > > > > >>What's your query like? Can you share it? Do you call any algebraic > > UDF > > > > >> after group by? I am wondering whether combiner matters in your > > test. > > > > I have been running 3 different types of queries. > > > > > > > > The first was performed on datasets of 6 different sizes: > > > > > > > > > > > > - Dataset size 1: 30,000 records (772KB) > > > > - Dataset size 2: 300,000 records (6.4MB) > > > > - Dataset size 3: 3,000,000 records (63MB) > > > > - Dataset size 4: 30 million records (628MB) > > > > - Dataset size 5: 300 million records (6.2GB) > > > > - Dataset size 6: 3 billion records (62GB) > > > > > > > > The datasets scale linearly, whereby the size equates to 3000 * 10n . > > > > A seventh dataset consisting of 1,000 records (23KB) was produced to > > > > perform join > > > > operations on. Its schema is as follows: > > > > name - string > > > > marks - integer > > > > gpa - float > > > > The data was generated using the generate data.pl perl script > > available > > > > for > > > > download > > > > from https://issues.apache.org/jira/browse/PIG-200 to produce the > > > > datasets. The results are as follows: > > > > > > > > > > > > * * * * * * *Set 1 * *Set 2** * *Set 3** > > * > > > > *Set > > > > 4** * *Set 5** * *Set 6* > > > > *Arithmetic** * 32.82* * 36.21* * 49.49* * 83.25* > > > > * > > > > 423.63* * 3900.78 > > > > *Filter 10%** * 32.94* * 34.32* * 44.56* * 66.68* > > > > * > > > > 295.59* * 2640.52 > > > > *Filter 90%** * 33.93* * 32.55* * 37.86* * 53.22* > > > > * > > > > 197.36* * 1657.37 > > > > *Group** * * *49.43* * 53.34* * 69.84* * > > 105.12* > > > > *497.61* * 4394.21 > > > > *Join** * * * 49.89* * 50.08* * 78.55* * > > > 150.39* > > > > *1045.34* *10258.19 > > > > *Averaged performance of arithmetic, join, group, order, distinct > > select > > > > and filter operations on six datasets using Pig. Scripts were > > configured > > > as > > > > to use 8 reduce and 11 map tasks.* > > > > > > > > > > > > > > > > * * * Set 1** * *Set 2** * *Set 3** > * > > > > *Set > > > > 4** * *Set 5** * *Set 6* > > > > *Arithmetic** * 32.84* * 37.33* * 72.55* * > 300.08 > > > > 2633.72 27821.19 > > > > *Filter 10% * 32.36* * 53.28* * 59.22* * 209.5* > > > * > > > > 1672.3* *18222.19 > > > > *Filter 90% * 31.23* * 32.68* * 36.8* * 69.55* > > > > * > > > > 331.88* *3320.59 > > > > *Group * * * 48.27* * 47.68* * 46.87* * > 53.66* > > > > *141.36* *1233.4 > > > > *Join * * * * *48.54* *56.86* * 104.6* * > > > 517.5* > > > > * 4388.34* * - > > > > *Distinct** * * *48.73* *53.28* * 72.54* * > > > 109.77* > > > > * - * * * * - > > > > *Averaged performance of arithmetic, join, group, distinct select and > > > > filter operations on six datasets using Hive. Scripts were configured > > as > > > to > > > > use 8 reduce and 11 map tasks.* > > > > > > > > (If you want to see the standard deviation, let me know). > > > > > > > > So, to summarize the results: Pig outperforms Hive, with the > exception > > of > > > > using *Group By*. > > > > > > > > The Pig scripts used for this benchmark are as follows: > > > > *Arithmetic* > > > > -- Generate with basic arithmetic > > > > A = load '$input/dataset_300000000' using PigStorage('\t') as (name, > > age, > > > > gpa) PARALLEL $reducers; > > > > B = foreach A generate age * gpa + 3, age/gpa - 1.5 PARALLEL > $reducers; > > > > store B into '$output/dataset_300000000_projection' using > PigStorage() > > > > PARALLEL $reducers; > > > > > > > > * > > > > * > > > > *Filter 10%* > > > > -- Filter that removes 10% of data > > > > A = load '$input/dataset_300000000' using PigStorage('\t') as (name, > > age, > > > > gpa) PARALLEL $reducers; > > > > B = filter A by gpa < '3.6' PARALLEL $reducers; > > > > store B into '$output/dataset_300000000_filter_10' using PigStorage() > > > > PARALLEL $reducers; > > > > > > > > > > > > *Filter 90%* > > > > -- Filter that removes 90% of data > > > > A = load '$input/dataset_300000000' using PigStorage('\t') as (name, > > age, > > > > gpa) PARALLEL $reducers; > > > > B = filter A by age < '25' PARALLEL $reducers; > > > > store B into '$output/dataset_300000000_filter_90' using PigStorage() > > > > PARALLEL $reducers; > > > > > > > > * > > > > * > > > > *Group* > > > > A = load '$input/dataset_300000000' using PigStorage('\t') as (name, > > age, > > > > gpa) PARALLEL $reducers; > > > > B = group A by name PARALLEL $reducers; > > > > C = foreach B generate flatten(group), COUNT(A.age) PARALLEL > $reducers; > > > > store C into '$output/dataset_300000000_group' using PigStorage() > > > PARALLEL > > > > $reducers; > > > > * > > > > * > > > > *Join* > > > > A = load '$input/dataset_300000000' using PigStorage('\t') as (name, > > age, > > > > gpa) PARALLEL $reducers; > > > > B = load '$input/dataset_join' using PigStorage('\t') as (name, age, > > gpa) > > > > PARALLEL $reducers; > > > > C = cogroup A by name inner, B by name inner PARALLEL $reducers; > > > > D = foreach C generate flatten(A), flatten(B) PARALLEL $reducers; > > > > store D into '$output/dataset_300000000_cogroup_big' using > PigStorage() > > > > PARALLEL $reducers; > > > > > > > > Similarly, here the Hive scripts: > > > > *Arithmetic* > > > > SELECT (dataset.age * dataset.gpa + 3) AS F1, > (dataset.age/dataset.gpa > > - > > > > 1.5) AS F2 > > > > FROM dataset > > > > WHERE dataset.gpa > 0; > > > > > > > > *Filter 10%* > > > > SELECT * > > > > FROM dataset > > > > WHERE dataset.gpa < 3.6; > > > > > > > > *Filter 90%* > > > > SELECT * > > > > FROM dataset > > > > WHERE dataset.age < 25; > > > > > > > > *Group* > > > > SELECT COUNT(dataset.age) > > > > FROM dataset > > > > GROUP BY dataset.name; > > > > > > > > *Join* > > > > SELECT * > > > > FROM dataset JOIN dataset_join > > > > ON dataset.name = dataset_join.name; > > > > > > > > I will re-run the benchmarks to see whether it is the reduce or map > > side > > > > that is slower and get back to you later today. > > > > > > > > The other two benchmarks were slightly different: I performed > > transitive > > > > self joins in which Pig outperformed Hive. However once I added a > Group > > > By, > > > > Hive began outperforming Pig. > > > > > > > > I also ran the TPC-H benchmarks and noticed that Hive (surprisingly) > > > > outperformed Pig. However what *seems* to cause the actual > performance > > > > difference is the heavy usage of the Group By operator in all but 3 > > TPC-H > > > > test scripts. > > > > > > > > Re-running the scripts whilst omitting the the grouping of data > > produces > > > > the expected results. For example, running script 3 > > > > (q3_shipping_priority.pig) whilst omitting the Group By operator > > > > significantly reduces the runtime (to 1278.49 seconds real time > runtime > > > or > > > > a total of 12,257,630ms CPU time). > > > > > > > > The fact that the Group By operator skews the TPC-H benchmark in > favour > > > of > > > > Apache Hive is supported by further experiments: as noted earlier a > > > > benchmark was carried out on a transitive self-join. The former took > > Pig > > > an > > > > average of 45.36 seconds (real time runtime) to execute; it took Hive > > > 56.73 > > > > seconds. The latter took Pig 157.97 and Hive 180.19 seconds (again, > on > > > > average). However adding the Group By operator to the scripts turned > > the > > > > tides: Pig is now significantly slower than Hive, requiring an > average > > of > > > > 278.15 seconds. Hive on the other hand required only 204.01 to > perform > > > the > > > > JOIN and GROUP operations. > > > > > > > > Real time runtime is measured using the time -p command. > > > > > > > > Best Regards, > > > > Benjamin > > > > > > > > > > > > > > > > On 20 August 2013 19:56, Cheolsoo Park <[email protected]> wrote: > > > > > > > > > Hi Benjarmin, > > > > > > > > > > Can you describe which step of group by is slow? Mapper side or > > reducer > > > > > side? > > > > > > > > > > What's your query like? Can you share it? Do you call any algebraic > > UDF > > > > > after group by? I am wondering whether combiner matters in your > test. > > > > > > > > > > Thanks, > > > > > Cheolsoo > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Aug 20, 2013 at 2:27 AM, Benjamin Jakobus < > > > > [email protected] > > > > > >wrote: > > > > > > > > > > > Hi all, > > > > > > > > > > > > After benchmarking Hive and Pig, I found that the Group By > operator > > > in > > > > > Pig > > > > > > is drastically slower that Hive's. I was wondering whether > anybody > > > has > > > > > > experienced the same? And whether people may have any tips for > > > > improving > > > > > > the performance of this operation? (Adding a DISTINCT as > suggested > > by > > > > an > > > > > > earlier post on here doesn't help. I am currently re-running the > > > > > benchmark > > > > > > with LZO compression enabled). > > > > > > > > > > > > Regards, > > > > > > Ben > > > > > > > > > > > > > > > > > > > > >
