More embarrassing microbenchmars for D's GC.

[Justin Johansson]

On 16/09/2010 9:53 PM, Steven Schveighoffer wrote:
> On Thu, 16 Sep 2010 06:41:14 -0400, Rounin
> <davidjo at student.matnat.uio.no> wrote:
>
>> Thank you for that advice. I'm using GDC because it's available from
>> Ubuntu
>> Linux's package system, whereas DMD currently is not. (And the .deb
>> posted on
>> digitalmars.com is only for i386.) Hopefully, D will gain more
>> popularity and more
>> up-to-date tools will be made available.
>
> I've heard that dmd is a bit tricky to set up on a 64-bit system (mine
> is 32-bit), but it will give more credence to your position to use the
> latest available tools.
>
>>
>> By the way, today I re-compiled the program with a "std.gc.enable;"
>> right before
>> the final "return 0" statement, and it still runs in 0.68 seconds. So
>> either those
>> objects aren't marked as needing garbage collection, or it's really
>> just an issue
>> of keeping the garbage collector from activating too quickly.
>
> The GC runs are not scheduled. Essentially, you are probably
> encountering a syndrome where the GC runs more than it should. This
> usually occurs when incrementally allocating large amounts of memory
> without dropping any of it.
>
> The GC colleciton runs when allocating memory cannot find any free
> memory. Most likely this is what's happening:
>
> allocate -> No free memory available -> run GC collection, get nothing
> back -> allocate more memory from OS -> allocate...
>
> Of course, the allocations will succeed for a bit after getting OS
> memory, but the GC collection is run way too often (as your profiling
> suggests).
>
> Disabling it at the beginning, then enabling it at the end means it's
> going to run at most once. Essentially, you take out the "run GC
> collection" step, everything else is the same. Given that 87% of the run
> time is spent in the collection cycle, it's pretty obvious why this
> reduced your runtime ;)
>
> -Steve

The redux of this discussion is surely that no matter how "good" the
semantics of a PL are, the proof of the pudding is in the GC.  Given
the more-or-less well-founded value vs reference semantics in D, it
might be useful to condider region-based memory management idioms and
integrate these idioms into the language.  Perhaps we might see this
in D99 as a measure to circumvent traditional GC issues.