On Tue, 6 Mar 2007, Richard Wilson wrote:
> Dear Misc,
>
> I'm currently putting together a new load balancer for my company's web farm,
> using OBSD for CARP redundancy and stability. I've chosen Pound[0] as it seems
> to be very simple and fast, and I like what I perceive to be their somewhat
> OBSD-like philosophies of easily readable, easily checkable, reliable code.
>
> That's the intro, now the question: The Pound documentation recomends that I
> should use PCRE and the Google PerfTools tcmalloc to increase speed. I'm
> already having to compile OpenSSL for thread support, and pcre is supplied as
> a package, but I'm wary of tcmalloc. It doesn't appear to be available as a
> package or port, and it strikes me that a memory allocator is a pretty
> fundamental thing to be mucking around with. The developers spent some time
> getting OpenBSD's core libraries just so, and unless I've got a good reason
> I'm inclined to trust them. However, 50ns versus 300ns for a malloc/free[1]
> seems like it could be a pretty good reason, but our current load balancer
> seems to be doing ok for now without it.
>
> I dunno. Am I being overly paranoid, or should I stick with nice dependable
> old-fashioned malloc?
>
> Looking for Clue...
>
> Dave
>
> [0] http://www.apsis.ch/pound/
> [1] http://goog-perftools.sourceforge.net/doc/tcmalloc.html
There's more to malloc than speed. Our malloc has built-in features
to catch several programming errors and make heap based attacks more
difficult.
I see nothing of the sort for tcmalloc.
-Otto
