Re: [MM] Make mm counters per cpu instead of atomic V2

["KAMEZAWA Hiroyuki" <kamezawa.hiroyu@jp.fujitsu.com>]

Christoph Lameter wrote:
> On Fri, 6 Nov 2009, KAMEZAWA Hiroyuki wrote:
>
>> BTW, can't we have single-thread-mode for this counter ?
>> Usual program's read-side will get much benefit.....
>
> Thanks for the measurements.
>
> A single thread mode would be good. Ideas on how to add that would be
> appreciated.
>

Maybe there are some ways....At brief thought....
==
struct usage_counter {
    long rss;
    long file;
}


struct mm_struct {
    ....
    atomic_long_t  rss;   /* only updated when usage_counter is NULL */
    atomic_long_t  file;  /* only updated when usage_counter is NULL */
    struct usage_counter *usage;  /* percpu counter used when
                                     multi-threaded */
    .....
}

And allocate mm->usage only when the first CLONE_THREAD is specified.

if (mm->usage)
    access per cpu
else
    atomic_long_xxx

and read operation will be

    val = atomic_read(mm->rss);
    if (mm->usage)
        for_each_possible_cpu()....
==
Does "if" seems too costly ?

If this idea is bad, I think moving mm_counter to task_struct from
mm_struct and doing slow-sync is an idea instead of percpu.

for example

struct task_struct {
    ....
    mm_counter_t temp_counter;
    ....
};

struct mm_struct {
    .....
    atomic_long_t rss;
    atomic_long_t file;
};

And adds temp_counter's value to mm_struct at some good point....before
sleep ?
kswapd and reclaim routine can update mm_struct's counter, directly.
Readers just read mm_struct's counter.

Thanks,
-Kame









--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>