Github user brkyvz commented on a diff in the pull request:
https://github.com/apache/spark/pull/3200#discussion_r23241236
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/linalg/distributed/BlockMatrix.scala
---
@@ -0,0 +1,223 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.mllib.linalg.distributed
+
+import breeze.linalg.{DenseMatrix => BDM}
+
+import org.apache.spark._
+import org.apache.spark.mllib.linalg._
+import org.apache.spark.mllib.rdd.RDDFunctions._
+import org.apache.spark.rdd.RDD
+import org.apache.spark.storage.StorageLevel
+import org.apache.spark.util.Utils
+
+/**
+ * A grid partitioner, which stores every block in a separate partition.
+ *
+ * @param numRowBlocks Number of blocks that form the rows of the matrix.
+ * @param numColBlocks Number of blocks that form the columns of the
matrix.
+ * @param rowsPerBlock Number of rows that make up each block.
+ * @param colsPerBlock Number of columns that make up each block.
+ */
+private[mllib] class GridPartitioner(
+ val numRowBlocks: Int,
+ val numColBlocks: Int,
+ val rowsPerBlock: Int,
+ val colsPerBlock: Int,
+ override val numPartitions: Int) extends Partitioner {
+
+ /**
+ * Returns the index of the partition the SubMatrix belongs to.
+ *
+ * @param key The key for the SubMatrix. Can be its position in the grid
(its column major index)
+ * or a tuple of three integers that are the final row index
after the multiplication,
+ * the index of the block to multiply with, and the final
column index after the
+ * multiplication.
+ * @return The index of the partition, which the SubMatrix belongs to.
+ */
+ override def getPartition(key: Any): Int = {
+ key match {
+ case (rowIndex: Int, colIndex: Int) =>
+ Utils.nonNegativeMod(rowIndex + colIndex * numRowBlocks,
numPartitions)
+ case (rowIndex: Int, innerIndex: Int, colIndex: Int) =>
+ Utils.nonNegativeMod(rowIndex + colIndex * numRowBlocks,
numPartitions)
+ case _ =>
+ throw new IllegalArgumentException("Unrecognized key")
+ }
+ }
+
+ /** Checks whether the partitioners have the same characteristics */
+ override def equals(obj: Any): Boolean = {
+ obj match {
+ case r: GridPartitioner =>
+ (this.numPartitions == r.numPartitions) && (this.rowsPerBlock ==
r.rowsPerBlock) &&
+ (this.colsPerBlock == r.colsPerBlock)
+ case _ =>
+ false
+ }
+ }
+}
+
+/**
+ * Represents a distributed matrix in blocks of local matrices.
+ *
+ * @param numRowBlocks Number of blocks that form the rows of this matrix
+ * @param numColBlocks Number of blocks that form the columns of this
matrix
+ * @param rdd The RDD of SubMatrices (local matrices) that form this matrix
+ */
+class BlockMatrix(
+ val numRowBlocks: Int,
+ val numColBlocks: Int,
+ val rdd: RDD[((Int, Int), Matrix)]) extends DistributedMatrix with
Logging {
--- End diff --
Hi @mbhynes. Great question. We did have such a class, it was called
`BlockPartition`, which later was renamed to `SubMatrix`. I did try what you
suggested, but here was the catch. Even though the partition id's match for
partitions, they didn't necessarily end at the same executors. Spark
distributes the partitions deterministically, so theoretically, they should end
at the same executors, but what happens is, an executor lags, "something"
happens, and the partitions then get jumbled around executors. We couldn't
(Spark doesn't) guarantee that these partitions end up on the same machine for
fault tolerance reasons (it's how the scheduler works). Therefore we needed to
have indices as above (which the class `SubMatrix` had). To ensure that we add
the correct blocks with each other, calling a `.join` was inevitable. Instead
of storing the index inside both `SubMatrix` and having at as a key, we decided
to just index it as above.
Then the question becomes: Can users properly index their matrices,
properly supply `((Int, Int), Matrix)`? To solve this, we will support
conversions from `CoordinateMatrix` and `IndexedRowMatrix`, which are
convenient storage methods, and have users call `.toBlockMatrix` from these
classes.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
