[ https://issues.apache.org/jira/browse/SYSTEMML-1772?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Matthias Boehm resolved SYSTEMML-1772. -------------------------------------- Resolution: Fixed Assignee: Matthias Boehm Fix Version/s: SystemML 1.0 > Perftest: MultiLogReg 100M x 1K, sparse fails with OOM > ------------------------------------------------------ > > Key: SYSTEMML-1772 > URL: https://issues.apache.org/jira/browse/SYSTEMML-1772 > Project: SystemML > Issue Type: Bug > Reporter: Matthias Boehm > Assignee: Matthias Boehm > Fix For: SystemML 1.0 > > > Our perftest MultiLogReg 100M x 1K, sparse fails with the following OOM when > ran with 20GB driver budget. > {code} > java.lang.OutOfMemoryError: GC overhead limit exceeded > 17/07/14 13:42:04 WARN hdfs.BlockReaderFactory: I/O error constructing remote > block reader. > java.io.EOFException: Premature EOF: no length prefix available > at > org.apache.hadoop.hdfs.protocolPB.PBHelper.vintPrefixed(PBHelper.java:2282) > at > org.apache.hadoop.hdfs.RemoteBlockReader2.newBlockReader(RemoteBlockReader2.java:423) > at > org.apache.hadoop.hdfs.BlockReaderFactory.getRemoteBlockReader(BlockReaderFactory.java:818) > at > org.apache.hadoop.hdfs.BlockReaderFactory.getRemoteBlockReaderFromTcp(BlockReaderFactory.java:697) > at > org.apache.hadoop.hdfs.BlockReaderFactory.build(BlockReaderFactory.java:355) > at > org.apache.hadoop.hdfs.DFSInputStream.blockSeekTo(DFSInputStream.java:673) > at > org.apache.hadoop.hdfs.DFSInputStream.readWithStrategy(DFSInputStream.java:882) > at org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:934) > at java.io.DataInputStream.readFully(DataInputStream.java:195) > at java.io.DataInputStream.readFully(DataInputStream.java:169) > at org.apache.hadoop.io.SequenceFile$Reader.init(SequenceFile.java:1915) > at > org.apache.hadoop.io.SequenceFile$Reader.initialize(SequenceFile.java:1880) > at > org.apache.hadoop.io.SequenceFile$Reader.<init>(SequenceFile.java:1829) > at > org.apache.hadoop.io.SequenceFile$Reader.<init>(SequenceFile.java:1843) > at > org.apache.sysml.runtime.io.ReaderBinaryBlockParallel$ReadFileTask.call(ReaderBinaryBlockParallel.java:150) > at java.util.concurrent.FutureTask.run(FutureTask.java:266) > at > java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142) > at > java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) > at java.lang.Thread.run(Thread.java:748) > {code} > Thanks for catching this issue [~acs_s]. The root cause can be seen in the > following HOP characteristics and the generated runtime plan which contains a > CP mmchain operation for hop 456 > {code} > 17/07/14 16:48:50 INFO recompile.Recompiler: EXPLAIN RECOMPILE > GENERIC (lines 207-208): > --(432) TRead X [100000000,1000,1000,1000,999978303] [0,0,23270 -> 23270MB], > SPARK > --(439) r(t) (432) [1000,100000000,1000,1000,999978303] [23270,0,11444 -> > 34714MB], SPARK > --(431) TRead P [100000000,2,1000,1000,200000000] [0,0,1526 -> 1526MB], CP > --(436) rix (431) [100000000,1,1000,1000,-1] [1526,0,763 -> 2289MB], CP > --(1276) u(sprop) (436) [100000000,1,1000,1000,-1] [763,0,763 -> 1526MB], CP > --(429) TRead ssX_V [1000,1,1000,1000,1000] [0,0,0 -> 0MB], CP > --(437) ba(+*) (432,429) [100000000,1,1000,1000,-1] [23270,0,763 -> 24033MB], > SPARK > --(1275) b(*) (1276,437) [100000000,1,1000,1000,-1] [1526,0,763 -> 2289MB], CP > --(456) ba(+*) (439,1275) [1000,1,1000,1000,-1] [12207,0,0 -> 12207MB], CP > --(457) TWrite HV (456) [1000,1,1000,1000,-1] [0,0,0 -> 0MB], CP > {code} > The final matrix multiplication for {{t(X) tmp}} fits in CP and satisfied the > mmchain pattern However, mmchain avoids the transpose (assuming that X must > fit into memory given that t(X) fits in memory). Given our MCSR and CSR > representations this is not necessarily true because there each row has a > certain sparse row overhead independent of the number of non-zeros. > We should consider this scenario during execution type selection and send the > entire pattern to SPARK in these cases which is anyway a good idea because > the first matrix multiplications is already in SPARK. If the additional > broadcast and blocksize constraints are met we compile a SPARK mmchain, > otherwise two subsequent SPARK matrix multiplications. -- This message was sent by Atlassian JIRA (v6.4.14#64029)