Re: [Sbcl-help] multi-threaded allocation appears to delay garbage collection

[james a. <jam...@se...>]

good morning;

On 2012-02-23, at 20:09 , Paul Khuong wrote:

> In article <A92...@se...>,
>  james anderson <jam...@se...> wrote:
>
>> good evening, all;
>>
>> it appears that memory which is allocated under "multi-threaded"
>> circumstances is not reliably released.
>> once a thread which has allocated the memory has been destroyed and
>> no reference to the thread remains, the memory is released.
> [...]
>>
>> ---------------------------------------------------------------------
>> ;;; queue memory test
>> ;;;
>> ;;; this demonstrates that memory which has been allocated in a  
>> new thread is
>> not released until the
>> ;;; thread itself has been released.
>
> Nope, that's just the conservative GC in action; the bulk  
> allocation and
> the GC logics is completely thread agnostic.  Arbitrary data on the
> stack (or in registers, more rarely) that look like pointers are  
> treated
> as potential strong references.
>
> That's usually not too much of an issue.  Queues built on top of  
> conses
> are a particularly bad weakness, though: one stray reference to an
> already-dequeued cons also references all the later conses, via the  
> cdr.
> That's why sb-queue clears all the fields of dequeued records.
>
> (defun dq ()
>   (sb-thread:wait-on-semaphore *queue-semaphore*)
>   (if (eq *queue-pointer* *queue-header*)
>     (error "released to an empty queue."))
>   (let ((value (cadr *queue-header*)))  ;; That cons should be  
> NILed out
>     (setf *queue-header* (cdr *queue-header*))
>     value))
>
> Paul Khuong

thank you for the suggestion. i applied it to the queue management  
and did observe some improvement.
it was not, however, sufficient to allow sufficient memory to be  
released to avoid heap exhaustion.
i added logic to resource the (potentially numerous and large)  
objects which pass through the queues and observe that room now  
displays a count for those objects which corresponds to the contents  
in their resource cache and that the heap is no longer exhausted.

best regards, from berlin,