Re: PATCH 2.3.26: kmalloc GFP_ZERO
Jeff Garzik (jgarzik@mandrakesoft.com)
Tue, 09 Nov 1999 16:36:25 -0500
Andrea Arcangeli wrote:
>
> On Tue, 9 Nov 1999 yodaiken@chelm.cs.nmt.edu wrote:
>
> >> Are you still talking about a cache of zeroed pages? Looks a bad idea to
> >> me.
> >
> >The numbers say otherwise.
>
> So you are benchmarking a machine that spend time in the idle task.
>
> If you benchmarked a kernel compile make sure to compile with:
>
> make MAKE='make -j4' -j4 bzImage
>
> Then the numbers will agree with me (if you won't hit swap). Also make
> sure to apply my SMP scheduler patch in the aa patches.
>
> >The idle task will build a cache of zeroed pages. If nothing is in the
> >cache, you pay the standard cost.
>
> You'll also pay the check to see if the pool is empty before doing the
> memset by hand. You'll take the slow path in the case worth to optimize.
If you add "kcalloc" instead of hacking "kmalloc" as I did, drivers
wanting zeroed pages make the conscious decision of calling another
function, with a potentially slower path.
Such a split would also make it easier to enumerate
architecture-dependent paths. For some architectures may choose the
default -- kmalloc+memset -- and other archs can choose a page-zeroing
idle task, all without affecting kmalloc/kmem_cache_alloc
--
Jeff Garzik | Just once, I wish we would encounter
Building 1024 | an alien menace that wasn't immune to
MandrakeSoft, Inc. | bullets. -- The Brigadier, "Dr. Who"
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/