Re: fork/exec performance problem?

[rikki cattermole via Digitalmars-d]

On 15/06/2016 9:59 AM, Johan Holmberg via Digitalmars-d wrote:
Hi!
I'm trying to write a simple D program to emulate "parallel -u -jN", ie.
running a number of commands in parallel to take advantage of a
multicore machine (I'm testing on a 24-core Ubuntu machine).

I have written almost equivalent programs in C++ and D, and hoped that
they should run equally fast. But the performance of the D version
degrades when the number of commands increase, and I don't understand
why. Maybe I'm using D incorrectly? Or is it the garbage collector that
kicks in (even if I hope I don't allocate much memory after the initial
setup)?

My first testcase consisted of a file with 85000 C/C++ compilation
commands, to be run 24 in parallel. Most source files are really small
(different modules in the runtime library of a C/C++ compiler for
embedded development, built in different flavors).

If I invoke the D program 9 times with around 10000 (85000/9 to be
exact) commands each time, it performs almost on par with the C++
version. But with all 85000 files in one invokation, the D version takes
1.5 times as long (6min 30s --> 10min).

My programs (C++ and D) are really simple:

1) read all commands from STDIN into an array in the program
2) iterate over the array and keep N programs running at all times
3) start new programs with "fork/exec"
4) wait for finished programs with "waitpid"

If I compare the start of a 85000-run and a 10000-run, the 85000-run is
slower right from the start. I don't understand why? The only difference
must be that 85000-run has allocated a bigger array.

My D program can be viewed at:


https://bitbucket.org/holmberg556/examples/src/79ef65e389346e9957c535b77201a829af9c62f2/parallel_exec/parallel_exec_dlang.d

Any help would be appreciated.

/Johan Holmberg

This is more appropriate for D.learn.

Few things, disable the GC and force a collect in that while loop.
Next you're allocating hugely.

I would recommend replacing commands variable with some form of 
'smarter' array. Basically allocating in blocks instead of just 
appending one at a time. I'm not sure Appender is quite what you want 
here so it will be home made so to speak.

My revised edition:

import std.conv;
import std.stdio;
import std.string;
import std.process;
import core.stdc.stdlib;
import core.sys.posix.unistd;
import core.sys.posix.sys.wait;

int process_start(string cmdline) {
    int pid = fork();
    if (pid == -1) {
        perror("fork");
        exit(1);
    }
    else if (pid == 0) {
      string[3] argv = ["sh", "-c", cmdline];
      execvp("sh", argv);
      _exit(126);
    }
    else {
        return pid;
    }
    assert(0);
}

void process_wait(out int pid, out int status) {
    pid = waitpid(0, &status, 0);
    if (pid == -1) {
        perror("waitpid");
        exit(1);
    }
}

int
main(string[] argv)
{
        import core.memory;
        GC.disable;
        
    int maxrunning = 1;
    if (argv.length > 1) {
        maxrunning = to!int(argv[1]);
    }
    bool verbose = (argv.length > 2);

    string command;
    string[] commands;
    foreach (line; stdin.byLine()) {
        commands ~= line.idup;
    }

    if (verbose) {
        writeln("#parallel = ", maxrunning);
        writeln("#commands = ", commands.length);
        stdout.flush();
    }

    int next = 0;
    int nrunning = 0;
    while (next < commands.length || nrunning > 0) {
        while (next < commands.length && nrunning < maxrunning) {
            process_start(commands[next]);
            next++;
            nrunning++;
        }
        int pid;
        int exitstatus;
        process_wait(pid, exitstatus);
        nrunning--;
        if (exitstatus != 0) {
            writeln("ERROR: ...");
            exit(1);
        }
                GC.collect;
    }
    return 0;
}