在golang中并行递归扫描树

I have a binary tree that accessing a node is relatively fast, with the exception of the leaves - they might be 100-1000 times slower. I have a recursive algorithm that I would like to implement in go (I am new to it).

Because I have to get to the leaves to get benefits from the parallelism I need to parallelize the execution higher in the tree. This though might result in millions of goroutines. Limiting this with semaphore does not seem the 'go' way- there is no such sync primitive. Another concern I have is how expensive is, in fact, a channel, should I use wait group instead.

My tree is abstract and the algorithm runs over it identifying the items by level and index.

// l3               0
//               /    \
// l2          0        1
//           /  \     /   \
// l1       0    1   2     3
//         / \  / \ / \   / \
// l0     0  1 2  3 4 5  6  7

For example, I can use such to compute sum of all items in a vector with the function:

func Sum(level, index int, items []int) int {
    if level == 0 {return items[index]}
    return Sum(level-1, index*2, items) + Sum(level-1, index*2+1, items)
}

What should be my approach? Can someone point me to a recursive tree multithreaded algorithm implemented in go?

It sounds like you need a worker pool. Here's an example I just wrote: https://play.golang.org/p/NRM0yyQi8X

package main

import (
    "fmt"
    "sync"
    "time"
)

type Leaf struct {
    // Whatever
}

func worker(i int, wg *sync.WaitGroup, in <-chan Leaf) {
    for leaf := range in {
        time.Sleep(time.Millisecond * 500)
        fmt.Printf("worker %d finished work: %#v
", i, leaf)
    }
    fmt.Printf("worker %d exiting
", i)
    wg.Done()
}

func main() {
    var jobQueue = make(chan Leaf)
    var numWorkers = 10
    // the waitgroup will allow us to wait for all the goroutines to finish at the end
    var wg = new(sync.WaitGroup)
    for i := 0; i < numWorkers; i++ {
        wg.Add(1)
        go worker(i, wg, jobQueue)
    }

    // enqueue work (this goes inside your tree traversal.)
    for i := 0; i < 100; i++ {
        jobQueue <- Leaf{}
    }

    // closing jobQueue will cause all goroutines to exit the loop on the channel.
    close(jobQueue)
    // Wait for all the goroutines to finish
    wg.Wait()
}

I strongly suggest reading this excellent blog post from top to bottom:

https://blog.golang.org/pipelines

It covers not only an example of exactly what you need (i.e. parallelized file-tree walk calculating MD5 file checksums), but much much more:

  • fan-in/fan-out channel techniques
  • parallelism
  • pipeline cancellations via done channels
  • pipeline error chaining via error channels
  • bounded parallelism

The last topic, bounded parallelism, is used to ensure 'walking' large-node directory-trees do not create excessive go-routines: bounded.go