Re: Passing a CoordinateMatrix to SystemML
Thanks Matthias - unfortunately I'm still running into an ArrayIndexOutOfBounds exception both in reading the file as IJV and when calling dataFrametoBinaryBlock. Just to confirm: I downloaded and compiled the latest version using: git clone https://github.com/apache/systemml cd systemml mvn clean package mvn -version Apache Maven 3.3.9 Maven home: /usr/share/maven Java version: 1.8.0_151, vendor: Oracle Corporation Java home: /usr/lib/jvm/java-8-oracle/jre Default locale: en_US, platform encoding: UTF-8 OS name: "linux", version: "4.4.0-103-generic", arch: "amd64", family: "unix" I have a simple driver script written in Scala which calls the API methods. I compile the script using SBT (version 1.0.4) and submit using spark-submit (spark version 2.2.0). Here's how I'm calling the methods: val x = spark.read.parquet(inputPath).select(featureNames) val mc = new MatrixCharacteristics(199563535L, 71403L, 1024, 1024, 2444225947L) // as far as I know 1024x1024 is default block size in sysml? println("Reading Direct") val xrdd = RDDConverterUtils.dataFrameToBinaryBlock(jsc, x, mc, false, true) xrdd.count here is the stacktrace from calling dataFrameToBinaryBlock: java.lang.ArrayIndexOutOfBoundsException: 0 at org.apache.sysml.runtime.matrix.data.SparseRowVector.append(SparseRowVector.java:196) at org.apache.sysml.runtime.matrix.data.SparseBlockMCSR.append(SparseBlockMCSR.java:267) at org.apache.sysml.runtime.matrix.data.MatrixBlock.appendValue(MatrixBlock.java:685) at org.apache.sysml.runtime.instructions.spark.utils.RDDConverterUtils$DataFrameToBinaryBlockFunction.call(RDDConverterUtils.java:1067) at org.apache.sysml.runtime.instructions.spark.utils.RDDConverterUtils$DataFrameToBinaryBlockFunction.call(RDDConverterUtils.java:999) at org.apache.spark.api.java.JavaRDDLike$$anonfun$fn$7$1.apply(JavaRDDLike.scala:186) at org.apache.spark.api.java.JavaRDDLike$$anonfun$fn$7$1.apply(JavaRDDLike.scala:186) at org.apache.spark.rdd.RDD$$anonfun$mapPartitions$1$$anonfun$apply$23.apply(RDD.scala:797) at org.apache.spark.rdd.RDD$$anonfun$mapPartitions$1$$anonfun$apply$23.apply(RDD.scala:797) at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:38) at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:323) at org.apache.spark.rdd.RDD.iterator(RDD.scala:287) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:96) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:53) at org.apache.spark.scheduler.Task.run(Task.scala:108) at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:335) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624) at java.lang.Thread.run(Thread.java:748) and here is the stacktrace from calling "read()" directly: java.lang.ArrayIndexOutOfBoundsException: 2 at org.apache.sysml.runtime.matrix.data.SparseBlockCOO.sort(SparseBlockCOO.java:399) at org.apache.sysml.runtime.matrix.data.MatrixBlock.mergeIntoSparse(MatrixBlock.java:1784) at org.apache.sysml.runtime.matrix.data.MatrixBlock.merge(MatrixBlock.java:1687) at org.apache.sysml.runtime.instructions.spark.utils.RDDAggregateUtils$MergeBlocksFunction.call(RDDAggregateUtils.java:627) at org.apache.sysml.runtime.instructions.spark.utils.RDDAggregateUtils$MergeBlocksFunction.call(RDDAggregateUtils.java:596) at org.apache.spark.api.java.JavaPairRDD$$anonfun$toScalaFunction2$1.apply(JavaPairRDD.scala:1037) at org.apache.spark.util.collection.ExternalSorter$$anonfun$5.apply(ExternalSorter.scala:189) at org.apache.spark.util.collection.ExternalSorter$$anonfun$5.apply(ExternalSorter.scala:188) at org.apache.spark.util.collection.AppendOnlyMap.changeValue(AppendOnlyMap.scala:150) at org.apache.spark.util.collection.SizeTrackingAppendOnlyMap.changeValue(SizeTrackingAppendOnlyMap.scala:32) at org.apache.spark.util.collection.ExternalSorter.insertAll(ExternalSorter.scala:194) at org.apache.spark.shuffle.sort.SortShuffleWriter.write(SortShuffleWriter.scala:63) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:96) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:53) at org.apache.spark.scheduler.Task.run(Task.scala:108) at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:335) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624) at java.lang.Thread.run(Thread.java:748) Best, Anthony On Sun, Dec 24, 2017 at 3:14 AM, Matthias Boehm wrote: > Thanks again for c
Re: Build server down
The slow down was wide spread. I logged INFRA-15716 . On Sun, Dec 24, 2017 at 8:22 AM, Glenn Weidner wrote: > It appears there's an issue with Jenkins service since I'm able to access > https://sparktc.ibmcloud.com/repo/latest/ but not > https://sparktc.ibmcloud.com/jenkins/. Unfortunately, I don't have direct > access to the machine to debug further. > > Regards, > Glenn > > > [image: Inactive hide details for Matthias Boehm ---12/24/2017 07:21:37 > AM---Hi all, could somebody at the STC please find out what's w]Matthias > Boehm ---12/24/2017 07:21:37 AM---Hi all, could somebody at the STC please > find out what's wrong with our build > > From: Matthias Boehm > To: dev@systemml.apache.org > Date: 12/24/2017 07:21 AM > Subject: Build server down > -- > > > > Hi all, > > could somebody at the STC please find out what's wrong with our build > server? Thanks and Happy Holidays. > > Regards, > Matthias > > > > >
Re: Build server down
It appears there's an issue with Jenkins service since I'm able to access https://sparktc.ibmcloud.com/repo/latest/ but not https://sparktc.ibmcloud.com/jenkins/. Unfortunately, I don't have direct access to the machine to debug further. Regards, Glenn From: Matthias Boehm To: dev@systemml.apache.org Date: 12/24/2017 07:21 AM Subject:Build server down Hi all, could somebody at the STC please find out what's wrong with our build server? Thanks and Happy Holidays. Regards, Matthias
Build server down
Hi all, could somebody at the STC please find out what's wrong with our build server? Thanks and Happy Holidays. Regards, Matthias
Re: Passing a CoordinateMatrix to SystemML
Thanks again for catching this issue Anthony - this IJV reblock issue with large ultra-sparse matrices is now fixed in master. It likely did not show up on the 1% sample because the data was small enough to read it directly into the driver. However, the dataFrameToBinaryBlock might be another issue that I could not reproduce yet, so it would be very helpful if you could give it another try. Thanks. Regards, Matthias On 12/24/2017 9:57 AM, Matthias Boehm wrote: Hi Anthony, thanks for helping to debug this issue. There are no limits other than the dimensions and number of non-zeros being of type long. It sounds more like an issues of converting special cases of ultra-sparse matrices. I'll try to reproduce this issue and give an update as soon as I know more. In the meantime, could you please (a) also provide the stacktrace of calling dataFrameToBinaryBlock with SystemML 1.0, and (b) try calling your IJV conversion script via spark submit to exclude that this issue is API-related? Thanks. Regards, Matthias On 12/24/2017 1:40 AM, Anthony Thomas wrote: Okay thanks for the suggestions - I upgraded to 1.0 and tried providing dimensions and blocksizes to dataFrameToBinaryBlock both without success. I additionally wrote out the matrix to hdfs in IJV format and am still getting the same error when calling "read()" directly in the DML. However, I created a 1% sample of the original data in IJV format and SystemML was able to read the smaller file without any issue. This would seem to suggest that either there is some corruption in the full file or I'm running into some limit. The matrix is on the larger side: 1.9e8 rows by 7e4 cols with 2.4e9 nonzero values, but this seems like it should be well within the limits of what SystemML/Spark can handle. I also checked for obvious data errors (file is not 1 indexed or contains blank lines). In case it's helpful, the stacktrace from reading the data from hdfs in IJV format is attached. Thanks again for your help - I really appreciate it. 00:24:18 WARN TaskSetManager: Lost task 30.0 in stage 0.0 (TID 126, 10.11.10.13, executor 0): java.lang.ArrayIndexOutOfBoundsException at java.lang.System.arraycopy(Native Method) at org.apache.sysml.runtime.matrix.data.SparseBlockCOO.shiftRightByN(SparseBlockCOO.java:594) at org.apache.sysml.runtime.matrix.data.SparseBlockCOO.set(SparseBlockCOO.java:323) at org.apache.sysml.runtime.matrix.data.MatrixBlock.mergeIntoSparse(MatrixBlock.java:1790) at org.apache.sysml.runtime.matrix.data.MatrixBlock.merge(MatrixBlock.java:1736) at org.apache.sysml.runtime.instructions.spark.utils.RDDAggregateUtils$MergeBlocksFunction.call(RDDAggregateUtils.java:627) at org.apache.sysml.runtime.instructions.spark.utils.RDDAggregateUtils$MergeBlocksFunction.call(RDDAggregateUtils.java:596) at org.apache.spark.api.java.JavaPairRDD$$anonfun$toScalaFunction2$1.apply(JavaPairRDD.scala:1037) at org.apache.spark.util.collection.ExternalSorter$$anonfun$5.apply(ExternalSorter.scala:189) at org.apache.spark.util.collection.ExternalSorter$$anonfun$5.apply(ExternalSorter.scala:188) at org.apache.spark.util.collection.AppendOnlyMap.changeValue(AppendOnlyMap.scala:150) at org.apache.spark.util.collection.SizeTrackingAppendOnlyMap.changeValue(SizeTrackingAppendOnlyMap.scala:32) at org.apache.spark.util.collection.ExternalSorter.insertAll(ExternalSorter.scala:194) at org.apache.spark.shuffle.sort.SortShuffleWriter.write(SortShuffleWriter.scala:63) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:96) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:53) at org.apache.spark.scheduler.Task.run(Task.scala:108) at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:335) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624) at java.lang.Thread.run(Thread.java:748) Anthony On Sat, Dec 23, 2017 at 4:27 AM, Matthias Boehm wrote: Given the line numbers from the stacktrace, it seems that you use a rather old version of SystemML. Hence, I would recommend to upgrade to SystemML 1.0 or at least 0.15 first. If the error persists or you're not able to upgrade, please try to call dataFrameToBinaryBlock with provided matrix characteristics of dimensions and blocksizes. The issue you've shown usually originates from incorrect meta data (e.g., negative number of columns or block sizes), which prevents the sparse rows from growing to the necessary sizes. Regards, Matthias On 12/22/2017 10:42 PM, Anthony Thomas wrote: Hi Matthias, Thanks for the help! In response to your questions: 1. Sorry - this was a typo: the correct schema is: [y: int, features: vector] - the column "features" was created using Spark's Vecto
Re: Passing a CoordinateMatrix to SystemML
Hi Anthony, thanks for helping to debug this issue. There are no limits other than the dimensions and number of non-zeros being of type long. It sounds more like an issues of converting special cases of ultra-sparse matrices. I'll try to reproduce this issue and give an update as soon as I know more. In the meantime, could you please (a) also provide the stacktrace of calling dataFrameToBinaryBlock with SystemML 1.0, and (b) try calling your IJV conversion script via spark submit to exclude that this issue is API-related? Thanks. Regards, Matthias On 12/24/2017 1:40 AM, Anthony Thomas wrote: Okay thanks for the suggestions - I upgraded to 1.0 and tried providing dimensions and blocksizes to dataFrameToBinaryBlock both without success. I additionally wrote out the matrix to hdfs in IJV format and am still getting the same error when calling "read()" directly in the DML. However, I created a 1% sample of the original data in IJV format and SystemML was able to read the smaller file without any issue. This would seem to suggest that either there is some corruption in the full file or I'm running into some limit. The matrix is on the larger side: 1.9e8 rows by 7e4 cols with 2.4e9 nonzero values, but this seems like it should be well within the limits of what SystemML/Spark can handle. I also checked for obvious data errors (file is not 1 indexed or contains blank lines). In case it's helpful, the stacktrace from reading the data from hdfs in IJV format is attached. Thanks again for your help - I really appreciate it. 00:24:18 WARN TaskSetManager: Lost task 30.0 in stage 0.0 (TID 126, 10.11.10.13, executor 0): java.lang.ArrayIndexOutOfBoundsException at java.lang.System.arraycopy(Native Method) at org.apache.sysml.runtime.matrix.data.SparseBlockCOO.shiftRightByN(SparseBlockCOO.java:594) at org.apache.sysml.runtime.matrix.data.SparseBlockCOO.set(SparseBlockCOO.java:323) at org.apache.sysml.runtime.matrix.data.MatrixBlock.mergeIntoSparse(MatrixBlock.java:1790) at org.apache.sysml.runtime.matrix.data.MatrixBlock.merge(MatrixBlock.java:1736) at org.apache.sysml.runtime.instructions.spark.utils.RDDAggregateUtils$MergeBlocksFunction.call(RDDAggregateUtils.java:627) at org.apache.sysml.runtime.instructions.spark.utils.RDDAggregateUtils$MergeBlocksFunction.call(RDDAggregateUtils.java:596) at org.apache.spark.api.java.JavaPairRDD$$anonfun$toScalaFunction2$1.apply(JavaPairRDD.scala:1037) at org.apache.spark.util.collection.ExternalSorter$$anonfun$5.apply(ExternalSorter.scala:189) at org.apache.spark.util.collection.ExternalSorter$$anonfun$5.apply(ExternalSorter.scala:188) at org.apache.spark.util.collection.AppendOnlyMap.changeValue(AppendOnlyMap.scala:150) at org.apache.spark.util.collection.SizeTrackingAppendOnlyMap.changeValue(SizeTrackingAppendOnlyMap.scala:32) at org.apache.spark.util.collection.ExternalSorter.insertAll(ExternalSorter.scala:194) at org.apache.spark.shuffle.sort.SortShuffleWriter.write(SortShuffleWriter.scala:63) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:96) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:53) at org.apache.spark.scheduler.Task.run(Task.scala:108) at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:335) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624) at java.lang.Thread.run(Thread.java:748) Anthony On Sat, Dec 23, 2017 at 4:27 AM, Matthias Boehm wrote: Given the line numbers from the stacktrace, it seems that you use a rather old version of SystemML. Hence, I would recommend to upgrade to SystemML 1.0 or at least 0.15 first. If the error persists or you're not able to upgrade, please try to call dataFrameToBinaryBlock with provided matrix characteristics of dimensions and blocksizes. The issue you've shown usually originates from incorrect meta data (e.g., negative number of columns or block sizes), which prevents the sparse rows from growing to the necessary sizes. Regards, Matthias On 12/22/2017 10:42 PM, Anthony Thomas wrote: Hi Matthias, Thanks for the help! In response to your questions: 1. Sorry - this was a typo: the correct schema is: [y: int, features: vector] - the column "features" was created using Spark's VectorAssembler and the underlying type is an org.apache.spark.ml.linalg.SparseVector. Calling x.schema results in: org.apache.spark.sql.types.StructType = StructType(StructField(features,org.apache.spark.ml. linalg.VectorUDT@3bfc3ba7,true) 2. "y" converts fine - it appears the only issue is with X. The script still crashes when running "print(sum(X))". The full stack trace is attached at the end of the message. 3. Unfortunately, the error persists when calling