R中的主/从多核处理

我想在以下代码中并行化while循环:

work <- function(n) {
  # Do some intensive work (e.g explore a graph starting at n).
  # After this, we don't need to execute work() on nodes in excluding.
  # (e.g exclude could be the nodes explored/reached from n)
  # n is just an example. exclude can be a potentially large set.
  Sys.sleep(2)
  exclude <- c(n, sample(nodes, rbinom(1, length(nodes), 0.5)))
  return(exclude)
}

nodes <- 1:1e3

#Order of execution doesn't matter
nodes <- sample(nodes)

#parallelize this loop
while(length(nodes) > 0) {
  n <- nodes[1]
  exclude <- work(n)
  nodes <- setdiff(nodes, exclude)
}

是否在排除节点上执行work()并不重要,但我们希望最小化此类实例.上面的while循环的目标是尽可能少地运行work()

这不是一个令人尴尬的并行计算,所以我不知道如何直接使用parLapply.可以使用主从框架,但我不知道任何多核编程(在Windows上).

作为具体示例,您可以将work(n)视为graph_exploration(n)(查找连接到n的所有节点的函数)并将其排除为n的连接组件中的节点.最终目标是从每个连接的组件中找到一个节点.您希望尽可能少地运行graph_exploration(n),因为这是一项昂贵的操作.

最佳答案 Miheer,

这是一个建议的解决方案.

前言:

这里的核心问题(据我所知)是解锁while循环,而work()是数字运算.基本上,只要资源保留以启动更多work()调用和处理,您就希望循环不被阻塞.好吧怎么办?好吧,我的建议是你使用future包.

下面的示例实际上为每个调用创建了一个work()的新进程调用.但是,除非所有已分配的工作进程都忙,否则调用不会阻止while循环.您可以看到这一点,因为每个work()调用都有一个不同的进程ID,如运行时输出中所示.

因此,每个作品()独立运行,以完成我们解决所有未来并返回决赛结果.

结果:

>顺序运行时间:经过20.61秒
>并行运行时间:经过8.22秒

我希望这能指出你正确的方向.

警告:您必须遍历所有节点,但它确实改善了运行时.

机器设置:

R version 3.4.1 (2017-06-30)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows >= 8 x64 (build 9200)
[Windows 10, 8 Core Xeon, 64Gb RAM]

并行代码示例:

# Check for, and install and load required packages.
requiredPackages <-
  c("tictoc", "listenv", "future")


ipak <- function(pkg) {
  new.pkg <- pkg[!(pkg %in% installed.packages()[, "Package"])]
  if (length(new.pkg))
    install.packages(new.pkg, dependencies = TRUE)
  sapply(pkg, require, character.only = TRUE)
}

ipak(requiredPackages)

work <- function(n) {
  # Do some intensive work (e.g explore a graph starting at n).
  # After this, we don't need to execute work() on nodes in exclude.
  # (e.g exclude could be the nodes explored/reached from n)
  # n is just an example. exclude can be a potentially large set.
  Sys.sleep(2) # sample(.5:5))
  exclude <- n
  return(exclude)
}

plan(multiprocess, workers = 4L)
#plan(sequential)

nodesGraph  <- 1:10
nodesGraph  <- sample(nodesGraph)
nodesCount  <- length(nodesGraph)
resultsList <- listenv()

tic()
while ( nodesCount > 0 ) {
  n <- nodesGraph[[nodesCount]]
  ## This is evaluated in parallel and will only block
  ## if all workers are busy.
  resultsList[[nodesCount]] %<-% {
      list( exclude = work(n), 
            iteration = length(nodesGraph), 
            pid = Sys.getpid())
  }

  nodesGraph <- setdiff(nodesGraph, nodesGraph[[nodesCount]] )
  cat("nodesGraph",nodesGraph,"\n")
  cat("nodesCount",nodesCount,"\n")
  nodesCount = nodesCount - 1
}
toc()

## Resolve all futures (blocks if not already finished)
resultsList <- as.list(resultsList)
str(resultsList)

并行运行时输出:

> source('<hidden>/dev/stackoverflow/47230384/47230384v5.R')
nodesGraph 2 5 8 4 6 10 7 1 9 
nodesCount 10 
nodesGraph 2 5 8 4 6 10 7 1 
nodesCount 9 
nodesGraph 2 5 8 4 6 10 7 
nodesCount 8 
nodesGraph 2 5 8 4 6 10 
nodesCount 7 
nodesGraph 2 5 8 4 6 
nodesCount 6 
nodesGraph 2 5 8 4 
nodesCount 5 
nodesGraph 2 5 8 
nodesCount 4 
nodesGraph 2 5 
nodesCount 3 
nodesGraph 2 
nodesCount 2 
nodesGraph  
nodesCount 1 
8.22 sec elapsed
List of 10
 $:List of 3
  ..$exclude  : int 2
  ..$iteration: int 1
  ..$pid      : int 10692
 $:List of 3
  ..$exclude  : int 5
  ..$iteration: int 2
  ..$pid      : int 2032
 $:List of 3
  ..$exclude  : int 8
  ..$iteration: int 3
  ..$pid      : int 16356
 $:List of 3
  ..$exclude  : int 4
  ..$iteration: int 4
  ..$pid      : int 7756
 $:List of 3
  ..$exclude  : int 6
  ..$iteration: int 5
  ..$pid      : int 10692
 $:List of 3
  ..$exclude  : int 10
  ..$iteration: int 6
  ..$pid      : int 2032
 $:List of 3
  ..$exclude  : int 7
  ..$iteration: int 7
  ..$pid      : int 16356
 $:List of 3
  ..$exclude  : int 1
  ..$iteration: int 8
  ..$pid      : int 7756
 $:List of 3
  ..$exclude  : int 9
  ..$iteration: int 9
  ..$pid      : int 10692
 $:List of 3
  ..$exclude  : int 3
  ..$iteration: int 10
  ..$pid      : int 2032

顺序运行时输出

> source('<hidden>/dev/stackoverflow/47230384/47230384v5.R')
nodesGraph 6 2 1 9 4 8 10 7 3 
nodesCount 10 
nodesGraph 6 2 1 9 4 8 10 7 
nodesCount 9 
nodesGraph 6 2 1 9 4 8 10 
nodesCount 8 
nodesGraph 6 2 1 9 4 8 
nodesCount 7 
nodesGraph 6 2 1 9 4 
nodesCount 6 
nodesGraph 6 2 1 9 
nodesCount 5 
nodesGraph 6 2 1 
nodesCount 4 
nodesGraph 6 2 
nodesCount 3 
nodesGraph 6 
nodesCount 2 
nodesGraph  
nodesCount 1 
20.61 sec elapsed
List of 10
 $:List of 3
  ..$exclude  : int 6
  ..$iteration: int 1
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 2
  ..$iteration: int 2
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 1
  ..$iteration: int 3
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 9
  ..$iteration: int 4
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 4
  ..$iteration: int 5
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 8
  ..$iteration: int 6
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 10
  ..$iteration: int 7
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 7
  ..$iteration: int 8
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 3
  ..$iteration: int 9
  ..$pid      : int 12484
 $:List of 3
  ..$exclude  : int 5
  ..$iteration: int 10
  ..$pid      : int 12484
点赞