Hi Bejoy,
Both my tables has 65m records ( ~ 1.8-1.9GB on hadoop) and bucketized on 'calling' column into 10 buckets. As you said, hive will load only 1 bucket ~ 180-190MB into memory. That's hardly to blow the heap (1.3GB) According to wiki, I set: set hive.input.format=org.apache.hadoop.hive.ql.io.BucketizedHiveInputFormat; set hive.optimize.bucketmapjoin = true; set hive.optimize.bucketmapjoin.sortedmerge = true; And run the following SQL select /*+ MAPJOIN(a) */ * from ra_md_cdr_ggsn_synthetic a join ra_ocs_cdr_ggsn_synthetic b on (a.calling = b.calling) where a.total_volume <> b.total_volume; But it still created many hash tables then threw Java Heap space error Best regards Nguyen Thanh Binh (Mr) Cell phone: (+84)98.226.0622 From: Bejoy Ks [mailto:bejoy...@yahoo.com] Sent: Thursday, April 05, 2012 3:07 PM To: user@hive.apache.org Subject: Re: Why BucketJoinMap consume too much memory Hi Amit Sorry for the delayed response, had a terrible schedule. AFAIK, there is no flags that would help you to take the hash table creation, compression and load into tmp files away from client node. From my understanding if you use a Map side join, the small table as a whole is converted into a hash table and compressed in a tmp file. Say if your child jvm size is 1gb and this small table is 5GB, it'd blow off jour job if the map tasks tries to get such a huge file in memory. Bucketed map join can help here, if the table is bucketed ,say 100 buckets then each bucket may have around 50mb of data. ie one tmp file would be just less that 50mb, here mapper needs to load only the required buckets in memory and thus hardly run into memory issues. Also on the client, The records are processed bucket by bucket and loaded into tmp files. So if your bucket size is too large, than the heap size specified for your client, it'd throw an out of memory. Regards Bejoy KS _____ From: Amit Sharma <amitsharma1...@gmail.com> To: user@hive.apache.org; Bejoy Ks <bejoy...@yahoo.com> Sent: Tuesday, April 3, 2012 11:06 PM Subject: Re: Why BucketJoinMap consume too much memory I am experiencing similar behavior in my queries. All the conditions for bucketed map join are met, and the only difference in execution when i set the hive.optimize.bucketmapjoin flag to true, is that instead of a single hash table, multiple hash tables are created. All the Hash Tables are still created on the client side and loaded into tmp files, which are then distributed to the mappers using distributed cache. Can i find any example anywhere, which shows behavior of bucketed map join, where in it does not create the has tables on the client itself? If so, is there a flag for it? Thanks, Amit On Sun, Apr 1, 2012 at 12:35 PM, Bejoy Ks <bejoy...@yahoo.com> wrote: Hi On a first look, it seems like map join is happening in your case other than bucketed map join. The following conditions need to hold for bucketed map join to work 1) Both the tables are bucketed on the join columns 2) The number of buckets in each table should be multiples of each other 3) Ensure that the table has enough number of buckets Note: If the data is large say 1TB(per table) and if you have just a few buckets say 100 buckets, each mapper may have to load 10GB>. This would definitely blow your jvm . Bottom line is ensure your mappers are not heavily loaded with the bucketed data distribution. Regards Bejoy.K.S _____ From: binhnt22 <binhn...@viettel.com.vn> To: user@hive.apache.org Sent: Saturday, March 31, 2012 6:46 AM Subject: Why BucketJoinMap consume too much memory I have 2 table, each has 6 million records and clustered into 10 buckets These tables are very simple with 1 key column and 1 value column, all I want is getting the key that exists in both table but different value. The normal did the trick, took only 141 secs. select * from ra_md_cdr_ggsn_synthetic a join ra_ocs_cdr_ggsn_synthetic b on (a.calling = b.calling) where a.total_volume <> b.total_volume; I tried to use bucket join map by setting: set hive.optimize.bucketmapjoin = true select /*+ MAPJOIN(a) */ * from ra_md_cdr_ggsn_synthetic a join ra_ocs_cdr_ggsn_synthetic b on (a.calling = b.calling) where a.total_volume <> b.total_volume; 2012-03-30 11:35:09 Starting to launch local task to process map join; maximum memory = 1398145024 2012-03-30 11:35:12 Processing rows: 200000 Hashtable size: 199999 Memory usage: 86646704 rate: 0.062 2012-03-30 11:35:15 Processing rows: 300000 Hashtable size: 299999 Memory usage: 128247464 rate: 0.092 2012-03-30 11:35:18 Processing rows: 400000 Hashtable size: 399999 Memory usage: 174041744 rate: 0.124 2012-03-30 11:35:21 Processing rows: 500000 Hashtable size: 499999 Memory usage: 214140840 rate: 0.153 2012-03-30 11:35:25 Processing rows: 600000 Hashtable size: 599999 Memory usage: 255181504 rate: 0.183 2012-03-30 11:35:29 Processing rows: 700000 Hashtable size: 699999 Memory usage: 296744320 rate: 0.212 2012-03-30 11:35:35 Processing rows: 800000 Hashtable size: 799999 Memory usage: 342538616 rate: 0.245 2012-03-30 11:35:38 Processing rows: 900000 Hashtable size: 899999 Memory usage: 384138552 rate: 0.275 2012-03-30 11:35:45 Processing rows: 1000000 Hashtable size: 999999 Memory usage: 425719576 rate: 0.304 2012-03-30 11:35:50 Processing rows: 1100000 Hashtable size: 1099999 Memory usage: 467319576 rate: 0.334 2012-03-30 11:35:56 Processing rows: 1200000 Hashtable size: 1199999 Memory usage: 508940504 rate: 0.364 2012-03-30 11:36:04 Processing rows: 1300000 Hashtable size: 1299999 Memory usage: 550521128 rate: 0.394 2012-03-30 11:36:09 Processing rows: 1400000 Hashtable size: 1399999 Memory usage: 592121128 rate: 0.424 2012-03-30 11:36:15 Processing rows: 1500000 Hashtable size: 1499999 Memory usage: 633720336 rate: 0.453 2012-03-30 11:36:22 Processing rows: 1600000 Hashtable size: 1599999 Memory usage: 692097568 rate: 0.495 2012-03-30 11:36:33 Processing rows: 1700000 Hashtable size: 1699999 Memory usage: 725308944 rate: 0.519 2012-03-30 11:36:40 Processing rows: 1800000 Hashtable size: 1799999 Memory usage: 766946424 rate: 0.549 2012-03-30 11:36:48 Processing rows: 1900000 Hashtable size: 1899999 Memory usage: 808527928 rate: 0.578 2012-03-30 11:36:55 Processing rows: 2000000 Hashtable size: 1999999 Memory usage: 850127928 rate: 0.608 2012-03-30 11:37:08 Processing rows: 2100000 Hashtable size: 2099999 Memory usage: 891708856 rate: 0.638 2012-03-30 11:37:16 Processing rows: 2200000 Hashtable size: 2199999 Memory usage: 933308856 rate: 0.668 2012-03-30 11:37:25 Processing rows: 2300000 Hashtable size: 2299999 Memory usage: 974908856 rate: 0.697 2012-03-30 11:37:34 Processing rows: 2400000 Hashtable size: 2399999 Memory usage: 1016529448 rate: 0.727 2012-03-30 11:37:43 Processing rows: 2500000 Hashtable size: 2499999 Memory usage: 1058129496 rate: 0.757 2012-03-30 11:37:58 Processing rows: 2600000 Hashtable size: 2599999 Memory usage: 1099708832 rate: 0.787 Exception in thread "Thread-1" java.lang.OutOfMemoryError: Java heap space My system has 4 PC, each has CPU E2180, 2GB ram, 80GB HDD, one of them containts NameNode, JobTracker, Hive Server and all of them contain DataNode, TaskTracker In all node, I set: export HADOOP_HEAPSIZE=1500 in hadoop-env.sh (~ 1.3GB heap) I want to ask you experts, why bucket join map consume too much memory? Am I wrong or my configuration is bad? Best regards,
