Re: [RFC][PATCH 00/18] Increase resolution of load weights

[Paul Turner]

On Thu, Apr 28, 2011 at 11:51 AM, Paul Turner <pjt@xxxxxxxxxx> wrote:
> On Thu, Apr 28, 2011 at 11:33 AM, Nikhil Rao <ncrao@xxxxxxxxxx> wrote:
>> On Thu, Apr 28, 2011 at 5:12 AM, Srivatsa Vaddagiri
>> <vatsa@xxxxxxxxxxxxxxxxxx> wrote:
>>> On Thu, Apr 28, 2011 at 05:18:27PM +0530, Nikunj A. Dadhania wrote:
>>>> --- kernel/sched.c.orig       2011-04-28 16:34:24.000000000 +0530
>>>> +++ kernel/sched.c    2011-04-28 16:36:29.000000000 +0530
>>>> @@ -1336,7 +1336,7 @@ calc_delta_mine(unsigned long delta_exec
>>>>                       lw->inv_weight = 1 + (WMULT_CONST - w/2) / (w + 1);
>>>>       }
>>>>
>>>> -     tmp = (u64)delta_exec * weight;
>>>> +     tmp = (u64)delta_exec * (weight >> SCHED_LOAD_RESOLUTION);
>>>
>>> Should we be fixing inv_weight rather to account for SCHED_LOAD_RESOLUTION here?
>>>
>>
>> Yes, I have been looking into fixing inv_weight and calc_delta_mine()
>> calculations based on the assumption that we have u64 weights. IMO the
>> function is complicated because the return value needs to be
>> calculated to fit into unsigned long. I would like to update users of
>> calc_delta_mine() to use u64 instead of unsigned longs and I think
>> this can be easily done (quick inspection of the code shows two call
>> sites that need to be updated - update_curr() and wakeup_gran()).
>> Without the restriction to fit into unsigned long, I think we can make
>> calc_delta_mine() and the inv_weight calculations simpler.
>>
>
> I don't think you have much room to maneuver here, the calculations in
> c_d_m() are already u64 based, even on 32bit.  Changing the external
> load factors to 64 bit doesn't change this.
>
> We lose fairness in cdm beyond 32 bits, at the old LOAD_SCALE=10
> you've got 22 bits with which you can maintain fairness. This gives
> total accuracy in total curr on any delta <= ~4ms (for a NICE_0 task).
>  If you bump this up (and don't downshift before computing the inverse
> as you are) then you start introducing rounding errors beyond ~4us.
>
> This would also be further exacerbated in sched_period() since that's
> using the total cfs_rq weight.
>

(Downshifting as you are nicely avoids all this because we don't
really need fairness at the load-balance resolution, as Nikunj points
out above it was just the overflow detection that was broken)

>> -Thanks,
>> Nikhil
>>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/