Re^5: How can I force thread switching?

[BrowserUk]

The actual code was simplified to caricature so that problems and bugs could occur rapidly.

Sorry. I should have said confusing to me rather than "confused". It was obviously artifacts of the reduction process that was leaving me in the dark as to the wider picture of your code.

The confused code may be a reference to my fiddling with is_running and is_joinable. This a (feeble?) attempt to protect against programming errors in the workers.... Since this is my first multi-thread Perl script, there may be a much better way to do it. Have you any suggestion?

The simplest approach to dealing with thread proc errors is to wrap them in block eval. Instead of:

my @threads = map threads->new( \&processrefsthread, $Q ), 1 .. $T;

Use:

sub threadProc {
    my $code = shift;
    for(1 .. 5 ){
        my @rv = eval { $code->( @_ );  };
        if( $@ ) {
            ## log errors if applicable
            ## next if retrying makes sense.
            ## last if not.
            tprint "$@; retrying";
        }
        else {
            return @rv; ## return any return values if applicable.
            ## think about scalar/list/void context.
        }
    }
}

my @threads = map threads->new( \&threadProc, \&processrefsthread, $Q 
+), 1 .. $T;

The retry loop is optional, but demonstrates the possibility. Adding that, plus a random divide-by-zero error in the actual thread proc:

sub processrefsthread {
    my $Q = shift;
    tprint 'Starting';
    while( my $job = $Q->dq() ) {
        tprint "processing job $job";

        my $d = int(rand(5));
        unless( $d ) {
            12345 / $d;
        }
        for my $i (0 .. $d * 500000) {
            my $j = $i * $d;
            $j = int(rand($j));
        }
    }
    tprint 'Ending';
}

And you get:

C:\test>1056140-buk
Useless use of division (/) in void context at C:\test\1056140-buk.p
[1] Starting
[2] Starting
[3] Starting
[4] Starting
[2] processing job 1
[1] processing job 2
[3] processing job 3
[4] processing job 4
[4] processing job 5
[4] processing job 6
[1] processing job 7
[3] processing job 8
[4] processing job 9
[4] Illegal division by zero at C:\test\1056140-buk.pl line 25.
; retrying
[4] Starting
[2] processing job 10
[4] processing job 11
[2] Illegal division by zero at C:\test\1056140-buk.pl line 25.
; retrying
[2] Starting
[2] processing job 12
[3] processing job 13
[2] processing job 14
[1] processing job 15
[1] Illegal division by zero at C:\test\1056140-buk.pl line 25.
; retrying
[1] Starting
...

Thus errors are trapped and diagnosed at source and recovery can be attempted without having to try and detect and divine reasoning remotely.

Concerning the logxxx, these are thread-safe (I hope) subs to write to the common screen.

A couple of thoughts:

• Rather than passing an arbitrary thread id into your threads and then passing that through to your log subs, use threads->tid to query the system provided, monotonically increasing tid within the log subs.
• In general, having in-place, scrolling messages on a screen that might be overwritten before they are noticed is historically a bad idea.

  Consider logging the verbose version of the messages to a file, and only very short status changes to the screen.

  Set and remember the highlight color of the most severe status for that job; and retain it even when subsequent status update the display. This way, even though the message the operator sees is a low priority 'done' message, the color (say red) tells them that something they perhaps missed occurred earlier and that they should investigate the log to find out what.

Your latest code shows you passed a reference to the queue as a thread argument. What is the advantage/difference compared to a :shared variable? What is the efficiency impact?

From an efficiency point of view, whether explicitly shared (via: my $Q :shared = ...), or implicitly cloned via closure:

my $Q - ...;

sub thread {
    while( $Q->dq( ) )
        ...
    }
}

or implicitly shared via argument passing as I showed, it makes little or no difference. In all three cases we are manipulating a reference to an explictly shared array (within the Q module), and the runtime effect is pretty much the same.

The main advantages of using the implicitly shared reference via argument passing is the same as preferring to pass arguments to subroutines rather than have them access global variables. Ie. scoping; action at a distance; visibility etc.

The advantages of my self-limiting queue implementation over your externally-limited implementation are:

1. The logic is self-contained inside the Q implementation.

   Neither the producer no consumer code needs know or consider anything about the queue size growth, thus they are simpler and cleaner to program.

2. Only one locking/syncing mechanism is required.

   This is the biggy!

   Inter-thread comms (ITC) mechanisms like queues already require a locking mechanism to protect their internals against simultaneous access.

   The underlying locking mechanism used by threads::shared and Thread::Semaphore (cond_vars) are very expensive. This is because every action -- set, clear, test or wait -- requires a call into the kernel. Calls into the kernel -- even if only to do an operation that doesn't require memory/instruction fencing or pipeline flushing, eg. testing a bit -- are very expensive because they involve ring 3 - ring 0 - ring 3 transitions which have been measured "to cost 1000-1500 cycles on most machines.".

   Using one locking mechanism is costly enough, but (in Perl/threads at least) unavoidable. But using 2 is far more than just twice as costly, because the two can interact badly to become very costly indeed. Understanding how this comes about is quite difficult to follow; very hard to explain; and impossible to demonstrate. Please either take my word for it; or do your own research :)

The upshot is that I've never used Thread::Semaphore; I had to install it to run your code, and I will be uninstalling it immediately. Not because it is badly implemented -- from a cursory glance it looks just fine -- but because there is nothing it can do for me that I cannot do better with the underlying locking mechanism applied directly to the shared resource to be controlled.

But if I use it, then I will have to use two separate locking mechanisms to control that single resource -- the use of threads::shared::lock() is unavoidable. And applying two locking mechanisms to a single resource is always a recipe for slow code at best; and dead/livelocks and priority inversions all too frequently.

I highly recommend that you drop your use of Thread::Semaphore and adopt my threads::Q implementation. The latter is now very well tested and proven and simplifies your use-case beyond measure.