Subject: Re: Thread benchmarks
To: matthew sporleder <email@example.com>
From: Andrew Doran <firstname.lastname@example.org>
Date: 09/28/2007 20:34:13
On Fri, Sep 28, 2007 at 01:57:58PM -0400, matthew sporleder wrote:
> Can you talk more about the malloc replacement?
There's a good bit of information at the URL below and the imlementation is
in FreeBSD's CVS. The main advantage to jemalloc is that it works well with
large numbers of threads.
Joerg has suggested what we try a few other BSD licensed allocators and see
what they are like, so it's not set in stone yet.
> Also- an interesting
> thing about benchmarks in the past was the long-running stability of
> netbsd. Did you see anything like that?
Well, what do you mean by stability? :-). The majority of the kinks have
been ironed out of the scheduler and thread library now, so the results on
NetBSD are constant given the same test setup and conditions. The one issue
that exists is that we are dropping a few TPS for every connection that's
added. NetBSD holds up like that until 900 simultaneous client threads. At
around 900 threads, some quite odd (and as yet unknown) behaviour is tickled
and the rate collapses to about 100tps.