Re: threads: work crew memory leak

[BrowserUk]

1. On Vista 64-bit / perl 5.10.1 / threads v1.76 / threads::shared v1.33 I see no signs of a memory leak.

   What versions of those modules do you have?

2. Having run for 100,000+ thread creation/deleteion cycles, the memory usage wobbles a bit but stays pretty much fixed around 35 or 35 MB mark. The occasional spikes just mean that at the instance that value was taken more threads had been created than destroyed in the last second or two. A situation that corrects itself immediately.

   An app that creates 100,000 threads in 15 minutes, each to handle 1 number, is really badly designed. Like building a new train for every journey to work. Inefficient and unsustainable.

3. All your shenanigans with a semaphore and detach and is_detached() are utterly redundant.

   Once you detach the thread, it ends. There is no possibility of attempting to "do it twice". And if you did attempt to detach the same thread twice, it does do no harm at all.

   This code functions identically with more clarity:

   #!/usr/bin/perl
   use strict;
   use warnings;
   use threads;
   use threads::shared;
   use Data::Dumper;
   
   $|=1;
   
   my $MAX_THREADS = 20;
   my $TERM :shared = 0;
   
   $SIG{'INT'} = $SIG{'TERM'} = sub {
           print("^C captured\n");
           $TERM=1;
   };
   
   sub stuff_thr($) {
       my ($job)=@_;
       my $tid=threads->tid();
   
       print "Hi, I am thread: $tid, I need to do something with $job\n";
   }
   
   sub main() {
       my @jobs = ( 1 .. 100 );
       
       while (@jobs && ! $TERM) {
   
           for( 1 .. $MAX_THREADS - threads->list() ) {
               my $job = shift(@jobs);
               last if (! $job);
   
               threads->create('stuff_thr',$job)->detach;
           }
   
           @jobs = ( 1 .. 100 ) if @jobs < 10
       }
       sleep 1 while threads->list() > 0;
   }
   
   main();
   

[rakzer]

I am using perl v5.12.2, threads v1.75, threads::shared v1.32.

My shenanigans with the semaphores and the detaching was a desperate attempt to plug the memory leak but of course you're right, it was useless. As others replied this could be an underlying problem with pthread. For now I'll rewrite the code to reuse threads.

[BrowserUk]

I am using perl v5.12.2, threads v1.75, threads::shared v1.32.

It might be worth your while trying threads v1.81 & threads::shared v1.34.

As others replied this could be an underlying problem with pthread.

*nix/pthreads isn't something I know much about, but my gut feel is this is more likely to be a problem with (perl) memory allocation than the underlying threading libraries.

One thing that should be done, if this is going to get fixed, is to find a minimal test case that demonstrates the problem and a bug report raised. Something like:

perl -Mthreads -wE"{sleep 1 while threads->list>20;async(sub{1})->deta
+ch;redo}"

Assuming that actually leaks on your system, because the memory usage is rock solid at 12.5MB here.

For now I'll rewrite the code to reuse threads.

Once you do, you'll never go back to using throw-away threads.

[zwon]

One thing that should be done, if this is going to get fixed, is to find a minimal test case that demonstrates the problem and a bug report raised.

It's already there, no need to clog RT with identical reports. #69598

[rakzer]

After 20 minutes the process used 2.3GB of memory. I will repeat the same test as soon as I have the proposed threads/threads::shared version available. Thank you very much for your help.

[perl-diddler]

An app that creates 100,000 threads in 15 minutes, each to handle 1 number, is really badly designed. Like building a new train for every journey to work. Inefficient and unsustainable.

In general (and in the specific case) I agree with you, however, threads are alot less useful on today's platforms due to their high overhead.

The idea of creating a separate thread for each instance of a non-progressive (each successive loop not dependent on previous) forloop to do each calculation on independent terms is horribly limited by the high overhead of threads.

Conceivably, EVEN a single addition of 100,000 variables on a 100,000 processor system could have a 100,000X speedup if thread creation overhead was 0.

Such a large gap between theory and practice...*sigh*....

[BrowserUk]

Conceivably, EVEN a single addition of 100,000 variables on a 100,000 processor system could have a 100,000X speedup if thread creation overhead was 0.

Sorry, but that's just very naive.

1. Firstly, even in C or assembler, the creation overhead can never be 0.

   The creation of a (kernel) thread requires, at minimum:

   • the allocation of a stack segment.
   • the allocation of a register set save area.
   • the allocation of a thread 'context' structure to hold stuff like priorities, permission etc.
   • linking that context into the scheduler dispatch queues and other control structures.

   And each of those requires a transition from ring3 user space to ring 0 kernel space, which costs about 800 clock cycles on its own.

2. Each time a thread runs, it requires that:
   • the current register contents be saved to that threads context structure;
   • the new threads saved register set be loaded into the registers;
   • the processor pipelines need to be flushed.
   • the scheduler queues and tables need to be updated.
   • it's almost inevitable that some L1/l2/l3 cache lines will need to be flushed to ram and reloaded.

   All of those will mean hundreds if not thousands of cycles overhead.

A single addition of 2 numbers happens in 1 clock cycle. Spawning a new thread for each addition of 100,000 numbers even if you had 100,000 cores on your processor--which isn't going to happen any time in the next 10 years, if at all--will take far longer than just looping over the whole 100,000 on a single thread. And that's in C, never mind an interpreted language like Perl.

However, if we get back into the real world and consider processing those 100,000 additions on a 4 or 16 or 64 core system, then starting 1 thread per core and having each operate on 100,000/cores values, it's quite likely that you can achieve real (and realistic) economies.

Of course, the best approach today and in the immediate future, would be to load the data onto a Tesla GPU, and let it's 448 cores loose on it in parallel.

threads are alot less useful on today's platforms due to their high overhead.

I'd love to know what platforms you're talking about where the cost of threading was cheaper?

The only ones I'm aware of that might qualify are things like Java 1.1's 'green' threads. Whilst they were cheap to spawn--as they run completely in user space, so avoiding ring-level transitions--they were practically useless in as much as they only emulate true multi-tasking, by embedding a scheduler within the VM. As such they don't scale across cores, so when any one green thread is running, all the others are stopped. So there is no performance gain for cpu-bound processing.

[rakzer]

FWIW, the reason why I originally chose to create that many threads is that threads could (and quite frequently will) time out due to network related issues. Creating a timer thread to kill timed out threads seemed very easy and handy. Creating a fork within a thread with a signal handler killing the fork on the other hand would be a lot more expensive.