Re: [PATCH 4/5] x86: Add config variables for SMP_MAX

[Ingo Molnar <mingo@elte.hu>]

* Mike Travis <travis@sgi.com> wrote:

> >> +config THREAD_ORDER
> >> +	int "Kernel stack size (in page order)"
> >> +	range 1 3
> >> +	depends on X86_64_SMP
> >> +	default "3" if X86_SMP_MAX
> >> +	default "1"
> >> +	help
> >> +	  Increases kernel stack size.
> >> +
> > 
> > Could you please elaborate, why this is needed and put more info 
> > about this requirement into this patch description?
> > 
> > People worked hard to push data allocation from stack to heap to 
> > make THREAD_ORDER of 0 and 1 possible. So why increase it again and 
> > why does this help scalability?
> > 
> > Many thanks and Best Regards
> > 
> > Ingo Oeser, puzzled a bit :-)
> 
> 
> The primary problem arises because of cpumask_t local variables.  
> Until I can deal with these, increasing NR_CPUS to a really large 
> value increases stack size dramatically.

those should be fixed:

> Here are the top stack consumers with NR_CPUS = 4k.
> 
>                          16392 isolated_cpu_setup
>                          10328 build_sched_domains
>                           8248 numa_initmem_init
>                           4664 cpu_attach_domain
>                           4104 show_shared_cpu_map
>                           3656 centrino_target
>                           3608 powernowk8_cpu_init
>                           3192 sched_domain_node_span
>                           3144 acpi_cpufreq_target
>                           2584 __svc_create_thread
>                           2568 cpu_idle_wait
>                           2136 netxen_nic_flash_print
>                           2104 powernowk8_target
>                           2088 _cpu_down
>                           2072 cache_add_dev
>                           2056 get_cur_freq
>                              0 acpi_processor_ffh_cstate_probe
>                           2056 microcode_write
>                              0 acpi_processor_get_throttling
>                           2048 check_supported_cpu

(and most of that is performance-uncritical.)

	Ingo