Question

I am currently developing a Scala framework for multi-process, parallel algorithms using MPJ-Express (i.e. a Java MPI-variant).

MPJ-Express, and basically all MPI-variants, work by starting many processes with the same program. Since I have no control over the processes at runtime (my program does not spawn processes), I can’t use standard unit testing frameworks due to the following reasons:

1. Not all processes will own a local copy of the results after execution (ideally, the results should be collected at an arbitrary root-process)

2. Lack of control over standard-output. It is not easily possible to get output only from a single process.

3. Lack of flow control. All processes must enter the same test at the same time to allow single-program multiple data algorithms.

Number 2 is mainly the problem, as 3 might work as expected and 1 can be fixed with an extra communication operation. Does anyone have practical experience, or know of better unit testing strategies for multi-process algorithms?

EDIT

Right now, I seem to be having luck with the following code using scalacheck:

package it.vigtig.thesis.collection.scalacheck

import java.io.OutputStream
import java.io.PrintStream

import org.scalacheck.Prop.forAll
import org.scalacheck.Properties

import it.vigtig.thesis.env.DistEnv.globalRank
import it.vigtig.thesis.env.DistEnv.parallelize

object CollectionCheck {

  def main(args: Array[String]) {

    parallelize(args) {

      if (globalRank > 0) {
        Console.setOut(new PrintStream(new OutputStream() {
          def write(b: Int) { //nop    
          }
        }))
      }

      ST.main(Array())
    }

  }

}

object ST extends Properties("String") {

  //  def println(a: String*) = gprintln(a)
  property("startsWith") = forAll((a: String, b: String) => (a + b).startsWith(a))

  property("concatenate") = forAll((a: String, b: String) =>
    (a + b).length > a.length && (a + b).length > b.length)

  property("substring") = forAll((a: String, b: String, c: String) =>
    (a + b + c).substring(a.length, a.length + b.length) == b)

}

The above code reroutes the scala-println method to a nop-operation for all other processes than p=0. I should be able to run the test-suites with parallel-methods, allowing only the root process to verify the results. The above yields the following output:

MPJ Express (0.38) is started in the multicore configuration
rank-0: 0.212437745 time taken for initialize
+ String.startsWith: OK, passed 100 tests.
! String.concatenate: Falsified after 0 passed tests.
> ARG_0: 
> ARG_1: 
+ String.substring: OK, passed 100 tests.

Answer 1

github.com/rickynils/scalacheck/issues/37 Apparently, it was due to a kind of bug in scalacheck. This bug will be fixed, but for now it seems you can clone the repository in that post and get a working copy! 🙂