Matthias Boehm created SYSTEMML-1772: ----------------------------------------
Summary: 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 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)