On Mon, Dec 14, 2009 at 10:26 AM, Justin Erenkrantz
<justin_at_erenkrantz.com> wrote:
> Just to note - in Amsterdam earlier this year, the APR/httpd folks
> ripped out pools and replaced it with a straight malloc/free
> implementation and the performance across the board was horrific even
> with the "best" malloc replacement libraries. So, that code was
> reverted...see:
>
> http://mail-archives.apache.org/mod_mbox/apr-dev/200903.mbox/%3c49CB671F.9040801@apache.org%3e
>
non-linux libc's:
http://mail-archives.apache.org/mod_mbox/apr-dev/200903.mbox/%3C4239a4320903260632ld52bd4h310853d66c157b5c@mail.gmail.com%3E
using tcmalloc:
http://mail-archives.apache.org/mod_mbox/apr-dev/200903.mbox/%3C4239a4320903260740l6786164l7d682c4d8a37d59b@mail.gmail.com%3E
The only way we could get it to approach or beat the existing APR
Pools was by using tcmalloc, which poses several problems for APR. In
addition, the variance of performance across platforms showed pretty
quickly that most libc's malloc is still 'pretty slow', although
things like FreeBSD's jemalloc are making the world better :)
The attempt to change back to native malloc/free was kicked off by
Ben's report. We had it in trunk for more than 2 months, we tried
optimizing it, but we just couldn't get it reliably into the same
level of performance.
Received on 2009-12-14 19:42:56 CET