Date: Wed, 13 Apr 2005 18:24:08 -0700 From: Julian Elischer <julian@elischer.org> To: Niall Douglas <s_sourceforge@nedprod.com> Cc: freebsd-threads@freebsd.org Subject: Re: thread aware malloc Message-ID: <425DC638.30801@elischer.org> In-Reply-To: <425E4810.7553.1133DEA0@localhost> References: <425DBC35.18797.F118E6C@localhost> <425E4810.7553.1133DEA0@localhost>
next in thread | previous in thread | raw e-mail | index | archive | help
Niall Douglas wrote: >On 14 Apr 2005 at 1:54, Ivan Voras wrote: > > > >>>KSE threads, FreeBSD v5.3. It's a mixture of process and system >>>threads. As I mentioned in a previous post, it's eight times slower >>>than Linux. See http://www.nedprod.com/Tn/ (scroll down to the >>>screenshots). >>> >>> >>Did you test on a "real" FreeBSD & Linux hardware (not VMWare)? >>Because VMWare **greatly** pessimizes low-level operations that depend >>on atomic/bus locks, CMPXCHG & similar operations that are used in >>synchronization, context switches & multithreading, and IO operations >>are also very very slow compared to real hardware. It simply cannot be >>used to do benchmarks on (except if you're benchmarking vmware, not >>the guest system :) ) >> >> > >VMWare should penalise Linux and FreeBSD equally in this area, >actually Linux even more as it doesn't have a CMPXCHG exemption from >kernel builds and Fedora Linux has more daemons etc. running by >default. > >Julian Elischer wrote: > > > >>did you compile the FreeBSD kernel with the required changes for running >>under vmware? >> >> > >Yes. It's a custom build made go as fast as I could make it. > >Interestingly, v5.3 was about half the speed of v5.2.1. > > > >>Vmware REALLY SUCKS when it comes to emulating the exact instructions we >>use for kernel locks and mutexes. You'd get maybe an order of magnitude >>difference through this under some situations. >> >>I forget the exact options but they'll be in the list archives. >>also make sure it's a Uniprocessor kernel. >> >> > >It's also a uniprocessor kernel. > >I finally have a native 64 bit platform which I just bought, I'm just >waiting on the release of v5.4. I'll post benchmarks here when I >know, but I'm still expecting an order of magnitude difference. >FreeBSD just feels a lot slower when doing multithreaded tasks >despite being faster for multiprocess work. It's as one would expect >with the added complexity of a M:N threading model which hasn't been >optimised yet. > > make sure you try libthr too. >Either way, ptmalloc2 is many times faster than the libc memory >allocator. It's over sixty times faster than the win32 default >allocator under four threads, so one would expect at least a similar >speed up. I would expect a six to twelve times speed up for real >world code. > >Cheers, >Niall > > > >_______________________________________________ >freebsd-threads@freebsd.org mailing list >http://lists.freebsd.org/mailman/listinfo/freebsd-threads >To unsubscribe, send any mail to "freebsd-threads-unsubscribe@freebsd.org" > >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?425DC638.30801>