I am moving from MPI based systems to Apache Spark. I need to do the following in Spark.
Suppose, I have n
vertices. I want to create an edge list from these n
vertices. An edge is just a tuple of two integers (u,v), no attributes are required.
However, I want to create them in parallel independently in each executor. Therefore, I want to create P
edge arrays independently for P
Spark Executors. Each array may be of different sizes and depends on the vertices, therefore, I also need the executor id from 0
to n-1
. Next, I want to have a global RDD Array of edges.
In MPI, I would create an array in each processor using the processor rank. How do I do that in Spark, especially using the GraphX
library?
Therefore, my primary goal is to create an array of edges in each executor and combine them into one single RDD.
I am first trying one modified version of the Erdos--Renyi model. As a parameter I only have the number of nodes n and a probability p.
Suppose, executor i
has to process nodes from 101
to 200
. For any node say, node 101
, it will create edges from 101
to 102 -- n
with probability p. After each executor creates the allocated edges, I would instantiate the GraphX EdgeRDD
and VertexRDD
. Therefore, my plan is to create the edge lists independently in each executor, and merge them into RDD
.
Lets start with some imports and variables which will be required for downstream processing:
import org.apache.spark._
import org.apache.spark.graphx._
import org.apache.spark.rdd.RDD
import scala.util.Random
import org.apache.spark.HashPartitioner
val nPartitions: Integer = ???
val n: Long = ???
val p: Double = ???
Next we'll need an RDD of seed IDs which can be used to generate edges. A naive way to handle this would be simply something like this:
sc.parallelize(0L to n)
Since number of the generated edges depends on the node id this approach would give a highly skewed load. We can do a little bit better with repartitioning:
sc.parallelize(0L to n)
.map((_, None))
.partitionBy(new HashPartitioner(nPartitions))
.keys
but much better approach is to start with empty RDD and generate ids in place. We'll need a small helper:
def genNodeIds(nPartitions: Int, n: Long)(i: Int) = {
(0L until n).filter(_ % nPartitions == i).toIterator
}
which can be used as follows:
val empty = sc.parallelize(Seq.empty[Int], nPartitions)
val ids = empty.mapPartitionsWithIndex((i, _) => genNodeIds(nPartitions, n)(i))
Just a quick sanity check (it is quite expensive so don't use it in production):
require(ids.distinct.count == n)
and we can generate actual edges using another helper:
def genEdgesForId(p: Double, n: Long, random: Random)(i: Long) = {
(i + 1 until n).filter(_ => random.nextDouble < p).map(j => Edge(i, j, ()))
}
def genEdgesForPartition(iter: Iterator[Long]) = {
// It could be an overkill but better safe than sorry
// Depending on your requirement it could worth to
// consider using commons-math
// https://commons.apache.org/proper/commons-math/userguide/random.html
val random = new Random(new java.security.SecureRandom())
iter.flatMap(genEdgesForId(p, n, random))
}
val edges = ids.mapPartitions(genEdgesForPartition)
Finally we can create a graph:
val graph = Graph.fromEdges(edges, ())
来源:https://stackoverflow.com/questions/34296588/creating-array-per-executor-in-spark-and-combine-into-rdd