Re: a question on threads

[BrowserUk]

Rather than starting a new thread each time you might be better using a pool of long running threads that trigger from a time taken from queue.

The number of threads required to ensure prompt service will depend upon the time they take to do what they have to do. If the time is always less than 1 second, then you could get away with a pool of 1, but if they might take longer, you'll need more:

#! perl -slw
use strict;
use Time::HiRes qw[ time sleep ];
use threads;
use Thread::Queue;

$| = 1;

our $N            ||= 10;
our $THREADS    ||= 2;
our $DELAY        ||= 2;

my $running : shared = 0;

sub t{
    $running++;
    my( $Q )= @_;
    my $tid = threads->self->tid;
    while( my $time = $Q->dequeue ) {
        sleep .1 until time() >= $time;
        print "$tid ($time) : ", time();
        sleep $DELAY;
    }
    $running--;
}

my $Q = new Thread::Queue;

threads->create( \&t, $Q )->detach for 1 .. $THREADS;

my $start = int time();
$Q->enqueue( $_ ) for $start + 2  .. $start + 2 + $N;

$Q->enqueue( undef ) for 1 .. $THREADS;

sleep 1 while $running;

__END__
P:\test>455833 -THREADS=2 -N=10 -DELAY=1
1 (1115781909) : 1115781909
2 (1115781910) : 1115781910.09375
1 (1115781911) : 1115781911.09375
2 (1115781912) : 1115781912.07813
1 (1115781913) : 1115781913.07813
2 (1115781914) : 1115781914.0625
1 (1115781915) : 1115781915.0625
2 (1115781916) : 1115781916.04688
1 (1115781917) : 1115781917.04688
2 (1115781918) : 1115781918.03125
1 (1115781919) : 1115781919.03125

P:\test>455833 -THREADS=2 -N=10 -DELAY=3
1 (1115781925) : 1115781925.07813
2 (1115781926) : 1115781926.0625
1 (1115781927) : 1115781928.07813
2 (1115781928) : 1115781929.0625
1 (1115781929) : 1115781931.07813
2 (1115781930) : 1115781932.0625
1 (1115781931) : 1115781934.07813
2 (1115781932) : 1115781935.0625
1 (1115781933) : 1115781937.07813
2 (1115781934) : 1115781938.0625
1 (1115781935) : 1115781940.07813

P:\test>455833 -THREADS=3 -N=10 -DELAY=3
1 (1115781957) : 1115781957.07813
2 (1115781958) : 1115781958.0625
3 (1115781959) : 1115781959.04688
1 (1115781960) : 1115781960.07813
2 (1115781961) : 1115781961.0625
3 (1115781962) : 1115781962.04688
1 (1115781963) : 1115781963.07813
2 (1115781964) : 1115781964.0625
3 (1115781965) : 1115781965.04688
1 (1115781966) : 1115781966.07813
2 (1115781967) : 1115781967.0625

Note how in the second run above, there were not enough threads running for the delay and so they gradually get further and further out of sync each timeslot. And also, how adding one more thread brings that back in the third run.

Also, you'll get better accuracy by sleeping for a smaller time period and checking whether the timeslot has arrived than sleeping for an absolute time. The smaller the sleep, the greater the accuracy, but it will still never be always exact (within some small descrepancy) because you have to wait for the scheduler to give the appropriate thread a timeslice.

Also, reducing the sleep below a certain minimum will greatly increase the cpu time spent polling. A tenth of a second works well on my system, giving an accuracy of < 1/10th whilst burning negligable amounts of cpu.