[SYSTEMDS-1772] Perftest: MultiLogReg 100M x 1K, sparse fails with OOM

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: SystemML 0.15
• Assignee: Matthias Boehm
• Reporter: Matthias Boehm
• Created: 15/Jul/17 00:30
• Updated: 09/Sep/17 02:19
• Resolved: 15/Jul/17 03:48

Description

Our perftest MultiLogReg 100M x 1K, sparse fails with the following OOM when ran with 20GB driver budget.

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)

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

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

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.