Scalable way to access every element of ConcurrentHashMap exactly once

Question

I have 32 machine threads and one ConcurrentHashMap map, which contains a lot of keys. Key has defined a public method visit

Answer 1

The solution I will eventually go for is an array of ConcurrentHashMaps instead of one ConcurrentHashMap. This is ad hoc, but seems to be relevant for my usecase. I don't care about the second step being slow as it doesn't affect my code's performance. The solution is:

Object Creation:

1. Create an array of size t of ConcurrentHashMaps, where t is a number of threads.
2. Create an array of Runnables, also of size t.

Array Population (single threaded, not an issue):

1. Create the keys and apply pre-hash function, which will return an int in the range 0 ... t-1. In my case simply modulo t.
2. Put the key in the hashmap, by accessing appropriate entry in the array. E.g. if the pre-hash has resulted in index 4, then go for hashArray[4].put(key)

Array Iteration (nicely multithreaded, performance gain):

1. Assign every thread from Runnables array a job of iterating over the hashmap with a corresponding index. This should give give a t times shorter iteration as opposed to single threaded.

To see the proof of concept code (as it's got some dependencies from the project I can't post it here) head towards my project on github

EDIT

Actually, implementing the above proof of concept for my system has proven to be time-consuming, bug-prone and grossly disappointing. Additionally I've discovered I would have missed many features of the standard library ConcurrentHashMap. The solution I have been exploring recently, which looks much less ad-hoc and much more promising is to use Scala, which produces bytecode that is fully interoperable with Java. The proof of concept relies on stunning library described in this paper and AFAIK it is currently IMPOSSIBLE to achieve a corresponding solution in vanilla Java without writing thousands lines of code, given the current state of the standard library and corresponding third-party libraries.

import scala.collection.parallel.mutable.ParHashMap

class Node(value: Int, id: Int){
    var v = value
    var i = id
    override def toString(): String = v toString
}

object testParHashMap{
    def visit(entry: Tuple2[Int, Node]){
        entry._2.v += 1
    }
    def main(args: Array[String]){
        val hm = new ParHashMap[Int, Node]()
        for (i <- 1 to 10){
            var node = new Node(0, i)
            hm.put(node.i, node)
        }

        println("========== BEFORE ==========")
        hm.foreach{println}

        hm.foreach{visit}

        println("========== AFTER ==========")
        hm.foreach{println}

    }
}

Answer 2

If I were you I'd just try iterating the key set of ConcurrentHashMap first. You could try passing the processing of keys off to a thread pool (in bundles, if the task is too light weight), or even to a ForkJoin task but you should do that only if it's really necessary.

Having said that you could use a ConcurrentSkipListMap, in which you can get a NavigableSet of keys. You can then take out partitions from this by using the subSet method. However, ConcurrentHashMap would have better performance for put, get operations (note also it would use compareTo rather than hashCode). Situations where this is better seems pretty unlikely.

Answer 3

I could try to inherit from ConcurrentHashMap, get my hands on the instances of its inner Segment, try to group them into 32 groups and work on each group separately. This sounds like a hardcore approach though.

Hardcore indeed but about the only thing I would see that would work. toArray() builds the array by doing an enumeration so no win there. I can't believe that a synchronized HashSet would be better unless the ratio of visit() runs to other map operations is pretty high.

The problem with the working with the Segments is that you are going to have to be extremely careful that your code is resilient because I assume other threads may be altering the table at the same time you are visiting the nodes and you need to avoid the inevitable race conditions. Delicate for sure.

The big question in my mind is if this is necessary? Has a profiler or timing runs shown to you that this is taking too long to visit() each of the keys in one thread? Have you tried to do a thread-pool for each visit() call and have one thread doing the enumeration and the pool threads doing the visit()?