Re^2: Win32 - flock() not FIFO?

[hennesse]

BrowserUK,

Thread Queues is a perfect solution to the logfile problem.

When you suggested using Thread Queues for my IPC before, I looked into them, but mistakenly thought they were like pipes – one thread to one other thread.

Now that I see the error of my ways, I’m considering using them for my IPC (or is it ITC?), but there’s one problem.

Three of my threads just sit around waiting for a datagram to arrive from another thread using select(,,,undef). The other two have to break out of the select every x seconds to do some housekeeping, so they use select(,,,timeout).

Unfortunately, dequeue() doesn’t have a timeout feature. I don’t see any way to use dequeue_nb since during periods of inactivity, it’d have to spin in a tight loop.

Any ideas?

Thanks - Dave

[BrowserUk]

Three of my threads just sit around waiting for a datagram to arrive from another thread using select(,,,undef). The other two have to break out of the select every x seconds to do some housekeeping, so they use select(,,,timeout).

Unfortunately, dequeue() doesn’t have a timeout feature. I don’t see any way to use dequeue_nb since during periods of inactivity, it’d have to spin in a tight loop....

Any ideas?

In general (and idealistically):Don't poll!

The really nice thing about threading is that you don't have to try and make one piece of code do more than one thing at a time. If you have two things to do -- eg. 1) wait for a message from somewhere; 2) periodically do some housekeeping; -- use two threads.

async {
    while( $Q->dequeue() ) {
        ## process message.
    }
};

async {
    until( $dieNow  ) {
        sleep $someTime;
        ## do housekeeping.
    }
};

In this way, you separate the concerns of the two. Each is free to get on with exactly what it needs to do at its earliest opportunity -- ie. its next time-slice --, and continue that single task to completion without ever concerning itself that it might be taking too long or the need to check what else needs to be done.

That there in a nutshell is what makes kernel threading superior to event-driven or cooperative time-sharing mechanisms. Anyone who has ever done a job that required concentration, and that they answer a phone in a timely fashion, will get this point completely. It can be done, but each task complicates, interferes with and detracts from the performance of the other.

</end-idealism>

If you were designing your application from scratch, I would definitely suggest that you remove the periodic housekeeping duties from the two 'wait-for-datagram' threads. And move them, if possible combining them, into their own thread who purpose is purely periodic housekeeping. Given that you are currently restructuring your application to use threading, I also strongly suggest that you at least consider this possibility at this point.

But, without having a good oversight of what your application is doing, or how it is currently structured and therefore what impact this type of re-structuring is likely to have, I have to realise that it may not be optimal of your time at this stage to do that.

So, you currently have something like:

while( my( $nfound, $timeleft) = select( ..., $timeout ) ) {
    if( $timeleft ) {
         ## Nothing doing, do some housekeeping
    }
    else {
        ## deal with the event
    }
}

Substitute:

use Time::HiRes qw[ sleep ];
...
    until( $dieNow ) {
        sleep( $timeout / 10 );
        if( my $dgram = $Q->dequeue_nb() ) {
             ## Act on $dgram
        }
        else {
             ## Do housekeeping (if required).
        }
    }

If your original timeout is 1 second, waking up every 1/10th of a second will have negligible affect on your CPU usage whilst ensuring that you process datagrams and housekeeping in a timely fashion. Indeed, you'd have to leave this:

async{ 1 while Win32::Sleep( 100 ); }->join;;

Running for a whole day to really notice that it had consumed any CPU at all.

[hennesse]

BrowserUK,

Threads / shared / queues opens up a lot of possibilities that will really simplify my application. Already I see how one of my select(...,timeout)'s will go away. The other one needs a little more thought, but I think it can be restructured along the lines you suggest.

Under the fork-exec model, the five processes start with static copies of the same variables. A change to any variable requires propagating it to the other processes. This is done with udp datagrams. Each of the four worker process talks only to the "supervisor", who relays the messages (see below).

There are three basic messages passed between processes:
Request: ("get", $item, "") or ("set", $item, $state)
Response: ( "status", $item, $state)

Four of the processes do actual work, while the "supervisor" primarily just relays messages. The get and set messages are one-to-one, but the status message must be cc'd to the other processes so everyone knows the current status of $item. To cut down on the number of messages, the supervisor doesn't cc anyone if the status has not changed. The supervisor also provides the "watchdog" function, waking up every 60 seconds. An external digital i/o PLC will do an "automated scram" if it has not received a command from the software in 120 seconds, so the watchdog function keeps it running. (Yes, the operator can initiate a "manual scram" too.) The watchdog also requests the status of any physical hardware device if no one else has asked for it in the last 60 seconds.

Enter threads / shared / queues: If I make %status a shared variable, then all the status response messages go away, since each thread will have access to the current $status of any item. Supervisor's job is now reduced to just doing the watchdog function: sleep(60); do_watchdog(); No select needed.

The gets and sets are unidirectional one-to-one messages, so each worker thread can have an incoming queue which it monitors with Q->dequeue. A thread can issue a get or set request directly to the receiving thread by doing a Q->enqueue. Now all the udp messages go away.

The other select(...,timeout) process needs some more thought. It must asynchronously process get and set requests, while waking up every second to implement the PID (proportional, integral, derivative) ramp-and-soak algorithm, and turn Solid State Relays on and off. As you suggest, waking up every 1/10 second to check the Q will not consume much CPU, and 1/10 sec is probably asynchronous enough.

Question: Do the shared variables and queues need locking?

With a single CPU, I suspect the shared variable updates are atomic, but the Q operations might not be. With multiple or multi-core CPUs, things could possibly happen at exactly the same time, so problems might occur. Thoughts?

 

If it ain't broke, don't fix it. Well, I've already converted to threads, the system works perfectly and consumes 0% CPU. Implementing shared variables and queues will take some work, but the benefit would be easier maintenance - there's a lot of complicated logic that would simply go away.

I've got some other questions that I'll post in new messages.

AH HA MOMENT: I got really tired of typing in all the P and /P tags to make my PerlMonks posts look pretty. Then I realized that if I composed them in Dreamweaver and typed in the "display" area, DW would make a new paragraph every time I hit the enter key. Much easier!

Dave
The Electronic Brewery

[BrowserUk]

If I make %status a shared variable, then all the status response messages go away,

Now you're cookin' on gas :)

Question: Do the shared variables and queues need locking?

Shared variables do. Queues don't.

(Or rather, they do as they are actually just shared arrays under the covers, but all the required locking is taken care of under those same covers, so you do not need to do it yourself.)

With regard to atomicity on a single-cpu system. Do not rely on it. Applying locks is very easy and not worth skipping for anything more important than trivial experiments.

there's a lot of complicated logic that would simply go away.

Indeed. The benefits of shared state and linear code flows.