Subject: Re: the path from nathanw_sa -> newlock2
To: Andrew Doran <firstname.lastname@example.org>
From: David Laight <email@example.com>
Date: 02/14/2007 00:19:17
On Tue, Feb 13, 2007 at 07:04:20PM +0000, Andrew Doran wrote:
> One or two of the advantages that SA provides, like reduced use of kernel
> stacks can be gotten in other ways. I'll be looking at improving the 1:1
> implementation as time goes on and I'm sure others will be too.
Also the kernel cost for a lwp is relatively small - probably < 16k on
i386 even with a kernel stack. So 1000 threads 'only' costs 16MB of kvm,
with a modern system, and a sane kva scheme, this probably doesn't matter.
Especially since the user stack space for the threads has to come out of
the 32bit user address space, and most of the kvm can be paged out.
The 1:1 threads also makes it much easier to use different schedulers
for different user threads (etc).
David Laight: firstname.lastname@example.org