RE: Re: [PATCH] sched/fair: Optimize CPU iteration using for_each_cpu_and[not]

[Li,Rongqing]

> On 15/08/25 09:15, lirongqing wrote:
> > From: Li RongQing <lirongqing@xxxxxxxxx>
> >
> > Replace open-coded CPU iteration patterns with more efficient
> > for_each_cpu_and() and for_each_cpu_andnot() macros in three locations.
> >
> > This change both simplifies the code and provides minor performance
> > improvements by using the more specialized iteration macros.
> >
> 
> TBF I'm not sure it does improve anything for the SMT cases considering we
> don't see much more than SMT8.
> 

I did the blow simple test on 128 cpu, smt 2 machine, and result shows for_each_cpu_andnot is better :

for_each_cpu + if()    vs  for_each_cpu_andnot()
5026373            vs   3398283
4034229            vs   2711302



#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/cpumask.h>
#include <linux/sched/clock.h>

static int test_init(void)
{

        int cpu, sibling;
        int i = 0;
        int loop = 1000;
        u64 now;

        now = local_clock();

        while (loop--) {
                for (cpu = 0; cpu < 128; cpu++) {
                        for_each_cpu(sibling, cpu_smt_mask(cpu)) {
                                if (cpu == sibling)
                                        continue;
                                i++;
                        }
                }
        }
        printk("%lld %d", local_clock() - now);

        i =0;
        loop = 1000;

        now = local_clock();
        while (loop--) {
                for (cpu = 0; cpu < 128; cpu++) {
                        for_each_cpu_andnot(sibling, cpu_smt_mask(cpu), cpumask_of(cpu)) {
                                i++;
                        }
                }
        }

        printk("%lld %d", local_clock() - now);


        return -1;
}

module_init(livepatch_init);
MODULE_LICENSE("GPL");
MODULE_INFO(livepatch, "Y");

Thanks

-Li





> The task_numa_find_cpu() one I do agree makes things better.
> 
> > Signed-off-by: Li RongQing <lirongqing@xxxxxxxxx>
> 
> Reviewed-by: Valentin Schneider <vschneid@xxxxxxxxxx>