Re: hot object memory patch

On Thu, Aug 10, 2000, Lincoln Dale wrote:
> At 02:07 10/08/00, Adrian Chadd wrote:
> >Just out of curiousity, has anyone actually tried bumping up SM_PAGE_SIZE
> >to say, 64kb, and seeing how that affects performance ?
>
> in a real-world environment, it'd probably make sense (and add to
> performance), provided the amount of physical ram is tuned accordingly.
>
> in terms of testing such a change with polygraph, i doubt it'd make any
> difference.
> we see <3% in-memory-hits with 2gb of in-memory "hot" objects with
> polygraph, when in a customer deployment, this is >55% of hits are served from.
>
> you need to have a _very_ good malloc library to handle the in-core memory
> allocations/deallocations.

Yup, that makes sense. I have a BSD kernel-like zone allocator floating
about which eliminates a lot of list walks whilst doing malloc operations,
but I won't be throwing that into the code for a while.

Adrian
Received on Thu Aug 10 2000 - 01:12:16 MDT