DragonFly kernel List (threaded) for 2011-01
Re: Turning on mpsafe mode by default in master
:> I'm going to remove the mplock shims in our global tokens in
:> master, essentially running the related code mpsafe from this point
:If you get the chance, can you benchmark the difference? I'm curious to
:see, say, the difference before and after for a buildworld/buildkernel.
I don't think buildworld/kernel would be effected much as those are
cpu-bound and (also) the link (ld) lines are serialized by the make
program, so they have reached their efficiency limit.
I do see a significant improvement when running something with a ton
of concurrency like blogbench. Namecache lookups can burst past
1 million lookups/second now and regularly clock in at 600-800K
The one exception is when running programs which fault a lot of pages.
A massively parallel compile, such as a bulk build, fits that bill and
processes wind up competing for the vm_token. This is still better than
competing for the mp_token, though.