On 8/25/05, Gavin Maltby <[EMAIL PROTECTED]> wrote:
> On 08/25/05 06:00, Felix Schulte wrote:
> > On 8/24/05, Eric Schrock <[EMAIL PROTECTED]> wrote:
> >> On Wed, Aug 24, 2005 at 10:47:31PM +0200, Felix Schulte wrote:
> >>> Why is Solaris x86 limited to 21 CPUs
> >>> (http://cvs.opensolaris.org/source/xref/usr/src/uts/i86pc/sys/machparam.h#51)?
> >>> Is there simply no hardware with more than 20 CPUs or are there other
> >>> limits in the kernel which cause this limitation (SPARC uses 64+576)?
> >> This is largely historical. There is a bug:
> >>
> >> 5081104 NCPU has been outgrown on x86
> > The problem is even far older -
> > http://bugs.opensolaris.org/bugdatabase/view_bug.do?bug_id=4238977 has
> > been filed in 1999 (six years ago).
> >
> > NCPU=64 for i386 would be nice to be in sync with SPARC.
>
> The change isn't trivial from a code or performance point of view.
> If you look at the sparc stuff you'll see that NCPU has a default but
> that can be overridden by specific platforms (that have more than
> the default, which I think is 32).
Its 64 for SPARC.
> We calculate the number of
> words required to represent the final NCPU value in a bitmask
> (ala cpuset_t) and define macros accordingly. Where we can fit
> in a single word the macros are trivial bit operations, but
> where we span several words (currently 9 for bigger sparc platforms,
> I think) we have to shuffle around words first. cpuset_t is used
> all over the show, so keeping it simple on platforms that do not
> need more than a single word is desirable.
Depends on how you define 'word' - nibble, byte, short, 32 bit, 64 bit
or 128 bit. 586 has 64bit datatypes, making the value 64 the optimum
for the near future (until machines with 16 sockets and quad-core cpus
or 8 sockets with two HT-capable dual cores become available).
--
_ Felix Schulte
_|_|_ mailto:[EMAIL PROTECTED]
(0 0)
ooO--(_)--Ooo
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
