On Friday 2011-01-21 01:43, David Miller wrote:
>From: Jan Engelhardt <jengelh@xxxxxxxxxx>
>Date: Fri, 21 Jan 2011 01:40:03 +0100 (CET)
>
>> Did I get something wrong? Is perhaps the Machine Description not
>> expressive enough to figure out how many packages (CPU sockets) there
>> are?
>
>This is simply how we number physical packages and "cores" on Niagara.
>
>CPU sockets are represented by NUMA nodes.
>
>We need three layers of scheduler grouping, and those are the three
>layers provided by the kernel's generic hierarchy:
>
> NUMA node --> physical_package --> core
>
>Inside the CPU socket we need two layers of scheduler grouping, so
>that's how I decided to implement things.
>
>So it's intentional and on purpose.
Linux 2.6.32 used to output the scheduler grouping on bootup (these
messages are gone in Linux 2.6.37):
CPU0 attaching sched-domain:
domain 0: span 0-3 level SIBLING
groups: 0 (cpu_power = 294) 1 (cpu_power = 294) 2 (cpu_power = 294) 3
(cpu_power = 294)
domain 1: span 0-3 level MC
groups: 0-3 (cpu_power = 1176)
domain 2: span 0-23 level CPU
groups: 0-3 (cpu_power = 1176) 4-7 (cpu_power = 1176) 8-11 (cpu_power = 1176) 12-15 (cpu_power = 1176) 16-19 (cpu_power = 1176) 20-23 (cpu_power = 1176)
SIBLING, MC, CPU. That looks pretty much like the three-level
grouping you described. Though if 0-23 makes up a CPU and 0-3
makes up a core (that's what one could infer), 0-3 is not
one thread.
I am still a bit puzzled since the T1 is configured quite like the
Intel i7 920:
T1: 4 threads à 6 cores à 1 CPU
core_sibling_list=0-3, thread_sibling_list=0-3
i7: 2 threads à 4 cores à 1 CPU
core_sibling_list=0-7, thread_sibling_list=0,4
And in comparison:
Altix4700: 2 threads à 2 cores à 128 CPUs
cpu0/core_sibling=0x03 (inferring>) core_sibling_list=0-3
cpu0/thread_sibling=0x01 (inferring>) thread_sibling_list=0-1
So it seems as if, because the toplogy is different between sparc
and {ia64, x86}, at least one has worse scheduling.
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
