Acceleration ETA algorithm

[phizel]

Was looking for a decent algorithm for determining the ETA of a long-running process, but everything on CPAN uses simplistic and inaccurate algorithms. Found this great article Benchmarking I/O ETA algorithms and converted the Acceleration algorithm to perl. And yes, it would be better to extract the state components into an object.

use Time::HiRes qw(time);
sub eta {
    my ($cur, $total, $time) = @_;
    return unless $cur and $time;

    state ($last_progress, $last_time, $last_v, $last_eta);
    state (@v, @eta, $window_size, $window_idx);
    state $init = do {
        ($last_progress, $last_time, $last_v, $last_eta) = (0, 0, 0, -
+1);
        ($window_size, $window_idx) = (10, 0);
    };
    state $sub_v_weight = sub { 1 + $_[0] };
    state $sub_eta_weight = sub { $_[0] ? 2 * $_[1] : 1 };
    state $sub_weighted_avg = sub {
        my ($sub_weight, $avg, $total_weight, $w) = (shift, 0, 0, 0);
        for my $i (0 .. $#_) {
            # first version messed up the index.
            my $j = ($i + @_ - $window_idx - 1) % @_;
            $w = $sub_weight->($j, $w);
            $avg += $w * $_[$i];
            $total_weight += $w;
        }
        return $avg / $total_weight;
    };

    my $v = ($cur - $last_progress) / (($time - $last_time) || 1);
    $v[$window_idx] = $v;
    $v = $sub_weighted_avg->($sub_v_weight, @v);

    if ($v and $last_v) {
        my ($min_v, $max_v) = $v < $last_v ? ($v, $last_v) : ($last_v,
+ $v);
        $v = $last_v + ($v - $last_v) * $min_v / $max_v;
    }

    my $a = ($v - $last_v) / ($last_time ? ($time - $last_time) : 1);
    my $r = $total - $cur;

    my $eta = $last_eta;
    if ($a and 0 < (my $d = ($v * $v + 2 * $a * $r))) {
        $eta = (sqrt($d) - $v) / $a;
    }
    elsif ($v) { $eta = $r / $v }

    $eta[$window_idx] = $eta;
    $eta = $sub_weighted_avg->($sub_eta_weight, @eta);

    ($last_progress, $last_time, $last_v, $last_eta, $window_idx)
        = ($cur, $time, $v, $eta, ($window_idx + 1) % $window_size);

    return $eta > 0 ? $eta : 0;
}

[Anonymous Monk]

How is this supposed to work?

say eta(1, 100, 1);

says 13.1067359796659 (shouldn't it be 100?). OK, with another call:

say eta(2, 100, 2);

says "Illegal division by zero" -- because $total_weight is zero.

[stevieb]

Could you include a common use case?

[phizel]

When you want to print out progress but don't actually want a progress bar like Term::ProgressBar.

[LanX]

> like Term::ProgressBar.

Hmm, the thing with progress bars (as I know them) is that they only move right. A considerable slow down will stall them, without giving a hint about a new prognosis.

They often only show the amount of work done, like percentage of downloaded data.

But an ETA can adapt and give a better feedback, how long to wait.

Like if a torrent download is losing connection to seeders, and the ratio drops.

I still have to read this test in detail, I'm wondering by which objective criteria those algorithms are judged. (?)

[LanX]

ETA = Estimated Time of Arrival

I had to search for it too. Don't think I heard it before in the context of applications.

The OP would have been clearer if it included at least one non-abbreviated reference.

It's akin to ongoing downloads estimating the remaining minutes till completion.

[LanX]

> And yes, it would be better to extract the state components into an object

I think it's easier (and better) rewritten in functional style with a generator function returning the ETA closure function.

Most of your states would be my-vars in the outer generators scope. And you wouldn't need the voodoo to initialize the first run.

This  state $init = do {... } is an interesting trick tho. :)