Re: [PATCH 1/4] sched,time: remove non-power-of-two divides from __acct_update_integrals

[Rik van Riel <riel@redhat.com>]

On 01/30/2016 09:44 AM, Frederic Weisbecker wrote:
> On Fri, Jan 29, 2016 at 10:36:02PM -0500, riel@redhat.com wrote:
>> From: Rik van Riel <riel@redhat.com>
>>
>> When running a microbenchmark calling an invalid syscall number
>> in a loop, on a nohz_full CPU, we spend a full 9% of our CPU
>> time in __acct_update_integrals.
>>
>> This function converts cputime_t to jiffies, to a timeval, only to
>> convert the timeval back to microseconds before discarding it.
>>
>> This patch leaves __acct_update_integrals functionally equivalent,
>> but speeds things up by about 12%, with 10 million calls to an
>> invalid syscall number dropping from 3.7 to 3.25 seconds.
>>
>> Signed-off-by: Rik van Riel <riel@redhat.com>
>> ---
>>  kernel/tsacct.c | 19 +++++++++----------
>>  1 file changed, 9 insertions(+), 10 deletions(-)
>>
>> diff --git a/kernel/tsacct.c b/kernel/tsacct.c
>> index 975cb49e32bf..41667b23dbd0 100644
>> --- a/kernel/tsacct.c
>> +++ b/kernel/tsacct.c
>> @@ -93,9 +93,9 @@ void xacct_add_tsk(struct taskstats *stats, struct task_struct *p)
>>  {
>>  	struct mm_struct *mm;
>>  
>> -	/* convert pages-usec to Mbyte-usec */
>> -	stats->coremem = p->acct_rss_mem1 * PAGE_SIZE / MB;
>> -	stats->virtmem = p->acct_vm_mem1 * PAGE_SIZE / MB;
>> +	/* convert pages-nsec/KB to Mbyte-usec, see __acct_update_integrals */
>> +	stats->coremem = p->acct_rss_mem1 * PAGE_SIZE / (1000 * KB);
>> +	stats->virtmem = p->acct_vm_mem1 * PAGE_SIZE / (1000 * KB);
>>  	mm = get_task_mm(p);
>>  	if (mm) {
>>  		/* adjust to KB unit */
>> @@ -125,22 +125,21 @@ static void __acct_update_integrals(struct task_struct *tsk,
>>  {
>>  	if (likely(tsk->mm)) {
>>  		cputime_t time, dtime;
>> -		struct timeval value;
>>  		unsigned long flags;
>>  		u64 delta;
>>  
>>  		local_irq_save(flags);
>>  		time = stime + utime;
>>  		dtime = time - tsk->acct_timexpd;
>> -		jiffies_to_timeval(cputime_to_jiffies(dtime), &value);
>> -		delta = value.tv_sec;
>> -		delta = delta * USEC_PER_SEC + value.tv_usec;
>> +		delta = cputime_to_nsecs(dtime);
> 
> You might want to add a comment specifying why we don't call cputime_to_usecs()
> directly (because we optimize if delta < TICK_NSEC).
> 
> Although this has a good impact on nohz_full, it might have a tiny bad one on !nohz_full
> because now we first convert jiffies to nsecs (which implies a multiplication by 1000)
> that we later divide again by 1000. Now this is ok because I plan to convert tsk->utime/stime
> to nsecs and thus remove most of the cputime_t use and conversions everywhere.

Isn't cputime_t in nanoseconds even on !nohz_full systems nowadays,
due to sched_clock?

Also, a multiplication is essentially instantaneous compared to
a division, which is why Peter suggested going this way around.

>>  
>> -		if (delta == 0)
>> +		if (delta < TICK_NSEC)
>>  			goto out;
> 
> 
>> +
>>  		tsk->acct_timexpd = time;
>> -		tsk->acct_rss_mem1 += delta * get_mm_rss(tsk->mm);
>> -		tsk->acct_vm_mem1 += delta * tsk->mm->total_vm;
>> +		/* The final unit will be Mbyte-usecs, see xacct_add_tsk */
>> +		tsk->acct_rss_mem1 += delta * get_mm_rss(tsk->mm) / 1024;
>> +		tsk->acct_vm_mem1 += delta * tsk->mm->total_vm / 1024;
> 
> The use of 1024 and the change on MB above are confusing me. Why are we doing that?
> 
> Thanks.

So the compiler can just do a right shift in the frequently called
code, and have no divide at all left in __acct_update_integrals.
However, reducing the value here does seem useful for the prevention
of overflows.

The divide is saved for when the statistics are read out to
userspace.

-- 
All rights reversed