cvs commit: src/sys/kern init_sysent.c syscalls.c systrace_args.c src/sys/sys syscall.h syscall.mk sysproto.h

Julian Elischer julian at elischer.org
Mon Mar 3 22:19:00 UTC 2008


Jeff Roberson wrote:
> 
> On Tue, 4 Mar 2008, Peter Jeremy wrote:
> 
>> On Mon, Mar 03, 2008 at 02:24:54AM -1000, Jeff Roberson wrote:
>>> it.  After all 1024 cpus should be enough forever.. ;)
>> 
>> I notice the smiley but...  Based on Sun's claims, they should
>> have 1024 thread boxes in 4-6 years.  They have 64-thread boxes
>> now (8 core x 8 thread) [*] and have said they'll double
>> threads/chip every 12-18 months.
>> 
>> [*] I thought there was a dual-chip T5xxx which would have 128
>> threads) but it doesn't show up.
> 
> We have much bigger problems than the CPU_SETSIZE in cpuset.
> Currently the kernel assumes it can fit all cpus in one register.
> So 64 is the max on any platform.  We need to modify the kernel to
> support this large number of cpus.  After that we'll still run
> poorly on most workloads with this many cpus due to internal lock
> contention.  There is a lot of work to do to support platforms of
> this size.

the whole UMA concept breaks down.
I don't know how we are going to cope when we need to
start breaking workloads upint osmall subunits and farm them out.

I think it's about 5 years out at most..

Cisco just announced their latest chip has 60 cores and
intel is playing with one with 80.. These cores are a little
like the CELL cores.. they not really designed to be used in a
UMA (symetrical) OS..  I heard rumours that microsoft is
diverting resources to sketch out an OS that uses these.
ironically they are in the best situation because they
can dedicate cores to do some of those
"render-green-box-on-frog-if-full-moon" type calls
they have.


> 
> 
>> 
>> -- Peter Jeremy Please excuse any delays as the result of my
>> ISP's inability to implement an MTA that is either
>> RFC2821-compliant or matches their claimed behaviour.
>> 



More information about the cvs-src mailing list