Thread overview
task parallelize dirEntries
Aug 11
Johnson
August 11
I've modified the sample from tour.dlang.org to calculate the md5 digest of the files in a directory using std.parallelism.

When I run this on a dir with huge number of files, I get:

core.exception.OutOfMemoryError@src/core/exception.d(696): Memory allocation failed

Since dirEntries returns a range, I thought std.parallelism.parallel can make use of that without loading the entire file list into the memory.

What am I doing wrong here? Is there a way to achieve what I'm expecting?

```
import std.digest.md;
import std.stdio: writeln;
import std.file;
import std.algorithm;
import std.parallelism;

void printUsage()
{
    writeln("Loops through a given directory and calculates the md5 digest of each file encountered.");
    writeln("Usage: md <dirname>");
}

void safePrint(T...)(T args)
{
    synchronized
    {
        import std.stdio : writeln;
        writeln(args);
    }
}

void main(string[] args)
{
    if (args.length != 2)
        return printUsage;

    foreach (d; parallel(dirEntries(args[1], SpanMode.depth).filter!(f => f.isFile), 1))
    {
        auto md5 = new MD5Digest();
        md5.reset();
        auto data = cast(const(ubyte)[]) read(d.name);
        md5.put(data);
        auto hash = md5.finish();
        import std.array;
        string[] t = split(d.name, '/');
        safePrint(toHexString!(LetterCase.lower)(hash), "  ", t[$-1]);
    }
}
```
August 11
On Friday, 11 August 2017 at 21:33:51 UTC, Arun Chandrasekaran wrote:
> I've modified the sample from tour.dlang.org to calculate the
>
> [...]

RHEL 7.2 64 bit
dmd v2.075.0
ldc 1.1.0
August 11
On Friday, 11 August 2017 at 21:33:51 UTC, Arun Chandrasekaran wrote:
> I've modified the sample from tour.dlang.org to calculate the md5 digest of the files in a directory using std.parallelism.
>
> When I run this on a dir with huge number of files, I get:
>
> core.exception.OutOfMemoryError@src/core/exception.d(696): Memory allocation failed
>
> Since dirEntries returns a range, I thought std.parallelism.parallel can make use of that without loading the entire file list into the memory.
>
> What am I doing wrong here? Is there a way to achieve what I'm expecting?
>
> ```
> import std.digest.md;
> import std.stdio: writeln;
> import std.file;
> import std.algorithm;
> import std.parallelism;
>
> void printUsage()
> {
>     writeln("Loops through a given directory and calculates the md5 digest of each file encountered.");
>     writeln("Usage: md <dirname>");
> }
>
> void safePrint(T...)(T args)
> {
>     synchronized
>     {
>         import std.stdio : writeln;
>         writeln(args);
>     }
> }
>
> void main(string[] args)
> {
>     if (args.length != 2)
>         return printUsage;
>
>     foreach (d; parallel(dirEntries(args[1], SpanMode.depth).filter!(f => f.isFile), 1))
>     {
>         auto md5 = new MD5Digest();
>         md5.reset();
>         auto data = cast(const(ubyte)[]) read(d.name);
>         md5.put(data);
>         auto hash = md5.finish();
>         import std.array;
>         string[] t = split(d.name, '/');
>         safePrint(toHexString!(LetterCase.lower)(hash), "  ", t[$-1]);
>     }
> }
> ```

Just a thought, maybe the GC isn't cleaning up quick enough? You are allocating and md5 digest each iteration.

Possibly, an opitimization is use use a collection of md5 hashes and reuse them. e.g., pre-allocate 100(you probably only need as many as the number of parallel loops going) and then attempt to resuse them. If all are in use, wait for a free one. Might require some synchronization.

August 12
On Friday, 11 August 2017 at 21:58:20 UTC, Johnson wrote:
> Just a thought, maybe the GC isn't cleaning up quick enough? You are allocating and md5 digest each iteration.
>
> Possibly, an opitimization is use use a collection of md5 hashes and reuse them. e.g., pre-allocate 100(you probably only need as many as the number of parallel loops going) and then attempt to resuse them. If all are in use, wait for a free one. Might require some synchronization.

John, thanks. That was it. md.d has nifty function that is straightforward than the OOP version.

```
void main(string[] args)
{
    foreach (d; parallel(dirEntries(args[1], SpanMode.depth).filter!(f => f.isFile), 1))
    {
        auto data = cast(const(ubyte)[]) read(d.name);
        auto hash = md5Of(data);
        import std.array;
        string[] t = split(d.name, '/');
        writeln(toHexString(hash), "  ", t[$-1]);
    }
}
```

Also I expected the performance to be faster than `md5sum`. However, that was not the case. Please see below. Is there anyway to optimize this further?

```
11-08-2017 17:22:54 vaalaham ~/code/d/d-mpmc-sample
$ time find /home/arun/downloads/boost_1_64_0/ -type f | xargs md5sum >/dev/null 2>&1

real    0m1.124s
user    0m0.952s
sys     0m0.208s
11-08-2017 17:23:16 vaalaham ~/code/d/d-mpmc-sample
$ ldc2 pmd.d -O3
11-08-2017 17:23:31 vaalaham ~/code/d/d-mpmc-sample
$ time ./pmd ~/downloads/boost_1_64_0 > /dev/null

real    0m0.499s
user    0m1.596s
sys     0m0.580s
11-08-2017 17:23:37 vaalaham ~/code/d/d-mpmc-sample
$
```

strace showed lots of futex exchanges. Why would that be?