Re: more xcalloc() -> mempool conversions!

On Mon, Sep 25, 2000, Andres Kroonmaa wrote:
> On 25 Sep 2000, at 17:52, Adrian Chadd <adrian@creative.net.au> wrote:
>
> >
> > This patch is a little rough but it converts a couple of "should have been
> > MemPool'ed a long time ago" structures.
> >
> > Can I get someone to look at this and tell me if its right? If so, I'll
> > commit it.
> >
>
> I wanted to ask few clarifications.
>
> 1. Why do we need to always clear alloced memory?
> memPool seems to only use calloc to actually get memory. When it gets
> memory from pools, it doesn't clear it, instead it clears mem always
> during free. Seems weird, although no difference.

Squid assumes that a newly-allocated "thing" is zeroed. Its bitten me
a few time. :)

> memset adds some overhead, even if very small, it adds up (especially
> on systems with limited CPU-ram bandwidth, like Intel). Do we actually
> need to clear the membuffers? Maybe there would be reason to split
> memPoolAlloc into separate versions: memPoolMalloc and memPoolCalloc?

Most of the time you'd be zeroing it anyway unless you were reading in some
socket buffer data.

> 2. memPool does not try to align allocations at any size. calloc() does
> it at small sizes (8). Maybe we should add a check that when allocations
> of VM page size granularity occurs, we should use valloc() or memalign()
> As Henrik has noted, this could speedup OS->userspace disk buffering by
> reducing mem-mem copying. Also, if we tune dlmalloc.c to use mmaps for
> allocations > 4-16KB, we could eventually somewhat reduce memory
> fragmentation also.
>

FreeBSD's malloc does page-align requests over a certain size. Feel free to
try optimising the memPool stuff. :-)

Remember that the MemPool code is very basic and there's a lot of room for
improvement.

Adrian

-- 
Adrian Chadd			"If a butterfly flaps its wings in China,
<adrian@creative.net.au>	    will a women get naked in Amsterdam?"
				      -- Ashley Penney on Chaos Theory