Python 多处理创建不正确的 pid

Question

我已经研究了一段时间了，但似乎无法弄清楚。我将它缩小到我的代码和操作系统(Linux 上的 Python 2.7.3)不同意应该运行哪些进程的情况。发生这种情况时，我的代码永远挂起，但不会抛出任何异常。有时代码会正常运行几个小时，有时只会运行几分钟，我想不通为什么。这表现如下。感谢您的观看，我真的很困惑(双关语)。

代码输出:

创建离散字符矩阵

running PoolWorker_82 (72 triplets), pid 25777, ppid 24892
running PoolWorker_83 (72 triplets), pid 25778, ppid 24892
running PoolWorker_84 (72 triplets), pid 25779, ppid 24892
running PoolWorker_85 (72 triplets), pid 25780, ppid 24892
running PoolWorker_86 (72 triplets), pid 25781, ppid 24892
running PoolWorker_87 (72 triplets), pid 25782, ppid 24892
running PoolWorker_88 (72 triplets), pid 25783, ppid 24892
running PoolWorker_89 (90 triplets), pid 25784, ppid 24892

ps aux 的输出...

1000     24892  2.0  0.9 559948 151088 pts/0   Sl+  09:14   0:16 p runsimulation.py
1000     25776  0.0  0.8 559932 138320 pts/0   S+   09:19   0:00 p runsimulation.py
1000     26015  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26021  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26023  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26025  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26027  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26029  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26031  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py
1000     26036  0.0  0.8 559948 138140 pts/0   S+   09:22   0:00 p runsimulation.py

可以看到父进程24982是有的，但是worker的pid没有。通常，这些会匹配，我可以看到工作人员 CPU 使用率在工作时达到 100%，然后在迭代完成后它们都消失了。当它失败时，我得到 pid 不匹配和使用 0.0% CPU 的进程(第 3 列)。

我的代码的相关部分如下(按调用顺序相反):

R 设置使用 rpy2 调用的函数:

def create_R(dir):
    """
    creates the r environment
    @param dir: the directory for the output files
    """
    r = robjects.r
    importr("phangorn")
    importr("picante")
    importr("MASS")
    importr("vegan")
    r("options(expressions=500000)")
    robjects.globalenv['outfile'] = os.path.abspath(os.path.join(dir, "trees.pdf"))
    r('pdf(file=outfile, onefile=T)')
    r("par(mfrow=c(2,3))")

    r("""
        generate_triplet = function(bits) {
        triplet = replicate(bits, rTraitDisc(tree, model="ER", k=2,states=0:1))
        triplet = t(apply(triplet, 1, as.numeric))
        sums = rowSums(triplet)
        if (length(which(sums==0)) > 0 && length(which(sums==3)) == 1) {
            return(triplet)
        }
        return(generate_triplet(bits))
        }
    """)

    r("""
        get_valid_triplets = function(numsamples, needed, bits) {
            tryCatch({
                m = generate_triplet(bits)
                while (ncol(m) < needed) {
                    m = cbind(m, generate_triplet(bits))
                }
            return(m)
            }, error = function(e){print(message(e))}, warning = function(e){print(message(e))})
        }
    """)

worker内部调用的函数:

def __get_valid_triplets(num_samples, num_triplets, bits, q):
    r = robjects.r
    name = current_process().name.replace("-", "_")
    timer = stopwatch.Timer()
    log("\trunning %s (%d triplets), pid %d, ppid %d" % (name, num_triplets, current_process().pid, os.getppid()),
        log_file)
    r('%s = get_valid_triplets(%d, %d, %d)' % (name, num_samples, num_triplets, bits))
    q.put((name, r[name]))
    timer.stop()
    log("\t%s complete (%s)" % (name, str(timer)), log_file)

设置池并使用 apply_async 调度工作程序的函数。工作人员写入托管队列，该队列在池加入后进行处理:

def __generate_candidate_discrete_matrix(num_cols, num_samples, sample_tree, bits, usable_cols):
    assert isinstance(sample_tree, dendropy.Tree)
    print "Creating discrete character matrix"
    r = robjects.r
    newick = sample_tree.as_newick_string()
    num_samples = len(sample_tree.leaf_nodes())
    robjects.globalenv['numcols'] = usable_cols
    robjects.globalenv['newick'] = newick + ";"
    r("tree = read.tree(text=newick)")
    r('m = matrix(nrow=length(tree$tip.label))') #create empty matrix
    r('m = m[,-1]') #drop the first NA column
    num_procs = mp.cpu_count()
    args = []
    div, mod = divmod(usable_cols, num_procs)
    [args.append(div) for i in range(num_procs)]
    args[-1] += mod
    for i, elem in enumerate(args):
        div, mod = divmod(elem, bits)
        args[-1] += mod
        args[i] -= mod
    manager = Manager()
    pool = Pool(processes=num_procs, maxtasksperchild=1)
    q = manager.Queue(maxsize=num_procs)
    for arg in args:
        pool.apply_async(__get_valid_triplets, (num_samples, arg, bits, q))
    pool.close()
    pool.join()

    while not q.empty():
        name, data = q.get()
        robjects.globalenv[name] = data
        r('m = cbind(m, %s)' % name)

    r('m = m[,1:%d]' % usable_cols)
    r('m = m[order(rownames(m)),]') # consistently order the rows 
    r('m = t(apply(m, 1, as.numeric))') # convert all factors given by rTraitDisc to numeric
    a = r['m']
    n = r('rownames(m)')
    return a, n

最后，第一个生成候选矩阵的函数被调用，确保它是一个有效的，如果不是，它将再次尝试使用一个新的矩阵。如果有效，它会在 R session 中存储一些东西并返回数据

def create_discrete_matrix(num_cols, num_samples, sample_tree, bits):
    """
    Creates a discrete char matrix from a tree
    @param num_cols: number of columns to create
    @param sample_tree: the tree
    @return: a r object of the matrix, and a list of the row names
    @rtype: tuple(robjects.Matrix, list)
    """
    r = robjects.r
    usable_cols = find_usable_length(num_cols, bits)
    a, n = __generate_candidate_discrete_matrix(num_cols, num_samples, sample_tree, bits, usable_cols)
    assert isinstance(a, robjects.Matrix)
    assert a.ncol == usable_cols

    paralin_matrix, valid = __create_paralin_matrix(a)
    if valid is False:
        sample_tree = create_tree(num_samples, type = "S")
        return create_discrete_matrix(num_cols, num_samples, sample_tree, bits)
    else:
        robjects.globalenv['paralin_matrix'] = paralin_matrix
        r('rownames(paralin_matrix) = rownames(m)')
        r('paralin_dist = as.dist(paralin_matrix, diag=T, upper=T)')
        r("paralinear_cluster = hclust(paralin_dist, method='average')")
    return sample_tree, a, n

Answer 1

看来这是通过服务器重启 (FML) 解决的。但是，获得了有效信息。将 worker 提交到池中时，请确保在 worker 本身中捕获异常，而不是在调用 pool.apply_async 的方法中捕获异常。

def __get_valid_triplets(num_samples, num_triplets, bits, q):
    try:
        r = robjects.r
        name = current_process().name.replace("-", "_")
        timer = stopwatch.Timer()
        log("\trunning %s (%d triplets), pid %d, ppid %d" % (name, num_triplets, current_process().pid, os.getppid()),
            log_file)
        r('%s = get_valid_triplets(%d, %d, %d)' % (name, num_samples, num_triplets, bits))
        q.put((name, r[name]))
        timer.stop()
        log("\t%s complete (%s)" % (name, str(timer)), log_file)
    except Exception, e:
        q.put("DEATH")
        traceback.print_exc()