Re: Proper (no) accounting for the disabled HTT cores

Maxim Sobolev wrote:
> Hi,
> 
> Currently, FreeBSD by default disables hyper-threading "cores", by not 
> scheduling any threads to it. However, it still counts those cores as 
> "active but permanently idle" when calculating system-wide CPUs 
> statistics. It is incorrect, since it skews statistics quite a bit and 
> creates real problems for certain types of applications (monitoring 
> applications for example), by making them believe that the system does 
> have enough idle resources, while in fact it does not.
> 
> I think the proper way to handle disabled cores is to not account for 
> them in any way. Please find the patch attached, which fixes the 
> problem. Any comments or suggestions are welcome.
> 
> -Maxim
>
> Index: local_apic.c
> ===================================================================
> RCS file: /home/ncvs/src/sys/i386/i386/local_apic.c,v
> retrieving revision 1.28
> diff -d -u -r1.28 local_apic.c
> --- local_apic.c    12 Jul 2006 21:22:43 -0000    1.28
> +++ local_apic.c    2 Sep 2006 00:42:32 -0000
> _at__at_ -615,6 +615,16 _at__at_
>      /* Send EOI first thing. */
>      lapic_eoi();
> 
> +    /*
> +     * Don't do any accounting for the disabled HTT cores, since it
> +     * will provide misleading numbers for the userland.
> +     *
> +     * No locking is necessary here, since even if we loose the race
> +     * when hlt_cpus_mask changes it is not a big deal, really.
> +     */
> +    if ((hlt_cpus_mask & (1 << PCPU_GET(cpuid))) != 0)
> +        return;
> +
>      /* Look up our local APIC structure for the tick counters. */
>      la = &lapics[PCPU_GET(apic_id)];
>      (*la->la_timer_count)++;

You probably also want this for amd64.

Also, wouldn't it be better to completely disable the LAPIC timer on the 
hyperthreaded cpus?

-- Suleiman